diff --git a/source/application/api/common/CMakeLists.txt b/source/application/api/common/CMakeLists.txt
new file mode 100644
index 0000000..5078adc
--- /dev/null
+++ b/source/application/api/common/CMakeLists.txt
@@ -0,0 +1,59 @@
+#----------------------------------------------------------------------------
+#  Copyright (c) 2022 Arm Limited. All rights reserved.
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#----------------------------------------------------------------------------
+
+#########################################################
+#  Common utility library used by use case libraries.   #
+#  NOTE: this library should not depend on HAL.         #
+#########################################################
+
+cmake_minimum_required(VERSION 3.15.6)
+
+set(COMMON_UC_UTILS_TARGET common_api)
+project(${COMMON_UC_UTILS_TARGET}
+    DESCRIPTION     "Common Utilities library"
+    LANGUAGES       CXX)
+
+# Create static library
+add_library(${COMMON_UC_UTILS_TARGET} STATIC)
+
+## Include directories - public
+target_include_directories(${COMMON_UC_UTILS_TARGET}
+    PUBLIC
+    include
+    ${TENSORFLOW_SRC_PATH}/tensorflow/lite/micro/tools/make/downloads/flatbuffers/include)
+
+## Sources
+target_sources(${COMMON_UC_UTILS_TARGET}
+    PRIVATE
+    source/Classifier.cc
+    source/ImageUtils.cc
+    source/Mfcc.cc
+    source/Model.cc
+    source/TensorFlowLiteMicro.cc)
+
+# Link time library targets:
+target_link_libraries(${COMMON_UC_UTILS_TARGET}
+    PUBLIC
+    log                     # Logging functions
+    arm_math                # Math functions
+    tensorflow-lite-micro)  # TensorFlow Lite Micro library
+
+# Display status:
+message(STATUS "*******************************************************")
+message(STATUS "Library                                : " ${COMMON_UC_UTILS_TARGET})
+message(STATUS "CMAKE_SYSTEM_PROCESSOR                 : " ${CMAKE_SYSTEM_PROCESSOR})
+message(STATUS "*******************************************************")
diff --git a/source/application/api/common/include/AudioUtils.hpp b/source/application/api/common/include/AudioUtils.hpp
new file mode 100644
index 0000000..cbf7bb7
--- /dev/null
+++ b/source/application/api/common/include/AudioUtils.hpp
@@ -0,0 +1,172 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef AUDIO_UTILS_HPP
+#define AUDIO_UTILS_HPP
+
+#include <cstddef>
+#include <cstdint>
+
+namespace arm {
+namespace app {
+namespace audio {
+
+    template<class T>
+    class SlidingWindow {
+    public:
+
+        /**
+         * @brief     Creates the window slider through the given data.
+         *
+         * @param[in] data         Pointer to the data to slide through.
+         * @param[in] dataSize     Size in T type elements wise.
+         * @param[in] windowSize   Sliding window size in T type wise elements.
+         * @param[in] stride       Stride size in T type wise elements.
+         */
+        SlidingWindow(T *data, size_t dataSize,
+                      size_t windowSize, size_t stride) {
+            m_start = data;
+            m_dataSize = dataSize;
+            m_size = windowSize;
+            m_stride = stride;
+        }
+
+        SlidingWindow() = default;
+
+        ~SlidingWindow() = default;
+
+        /**
+         * @brief  Get the next data window.
+         * @return Pointer to the next window, if next window is not available nullptr is returned.
+         */
+        virtual T *Next() {
+            if (HasNext()) {
+                m_count++;
+                return m_start + Index() * m_stride;
+            } else {
+                return nullptr;
+            }
+        }
+
+        /**
+         * @brief  Checks if the next data portion is available.
+         * @return true if next data portion is available.
+         */
+        virtual bool HasNext() {
+            return m_size + m_count * m_stride <= m_dataSize;
+        }
+
+        /**
+         * @brief Reset the slider to the initial position.
+         */
+        virtual void Reset() {
+            m_count = 0;
+        }
+
+        /**
+         * @brief     Resets the slider to the start of the new data.
+         *            New data size MUST be the same as the old one.
+         * @param[in] newStart   Pointer to the new data to slide through.
+         */
+        virtual void Reset(T *newStart) {
+            m_start = newStart;
+            Reset();
+        }
+
+        /**
+         * @brief  Gets current index of the sliding window.
+         * @return Current position of the sliding window in number of strides.
+         */
+        size_t Index() {
+            return m_count == 0? 0: m_count - 1;
+        }
+
+        /**
+         * @brief  Gets the index from the start of the data where the next window will begin.
+         *         While Index() returns the index of sliding window itself this function
+         *         returns the index of the data element itself.
+         * @return Index from the start of the data where the next sliding window will begin.
+         */
+        virtual uint32_t NextWindowStartIndex() {
+            return m_count == 0? 0: ((m_count) * m_stride);
+        }
+
+        /**
+         * @brief     Go to given sliding window index.
+         * @param[in] index   New position of the sliding window. If index is invalid
+         *                    (greater than possible range of strides) then next call to Next() will return nullptr.
+         */
+        void FastForward(size_t index) {
+            m_count = index;
+        }
+
+        /**
+         * @brief  Calculates whole number of times the window can stride through the given data.
+         * @return Maximum number of whole strides.
+         */
+         size_t TotalStrides() {
+            if (m_size > m_dataSize) {
+                return 0;
+            }
+            return ((m_dataSize - m_size)/m_stride);
+        }
+
+
+    protected:
+        T *m_start = nullptr;
+        size_t m_dataSize = 0;
+        size_t m_size = 0;
+        size_t m_stride = 0;
+        size_t m_count = 0;
+    };
+
+    /*
+     * Sliding window that will cover the whole length of the input, even if
+     * this means the last window is not a full window length.
+     */
+    template<class T>
+    class FractionalSlidingWindow : public SlidingWindow<T> {
+    public:
+        using SlidingWindow<T>::SlidingWindow;
+
+        /**
+         * @brief  Checks if the next data portion is available.
+         * @return true if next data portion is available.
+         */
+        bool HasNext() {
+            return this->m_count < 1 + this->FractionalTotalStrides() && (this->NextWindowStartIndex() < this->m_dataSize);
+        }
+
+        /**
+        * @brief  Calculates number of times the window can stride through the given data.
+        *         May not be a whole number.
+        * @return Number of strides to cover all data.
+        */
+        float FractionalTotalStrides() {
+            if (this->m_dataSize < this->m_size) {
+                return 0;
+            } else {
+                return ((this->m_dataSize - this->m_size) / static_cast<float>(this->m_stride));
+            }
+        }
+    };
+
+
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* AUDIO_UTILS_HPP */
\ No newline at end of file
diff --git a/source/application/api/common/include/BaseProcessing.hpp b/source/application/api/common/include/BaseProcessing.hpp
new file mode 100644
index 0000000..a54dd12
--- /dev/null
+++ b/source/application/api/common/include/BaseProcessing.hpp
@@ -0,0 +1,67 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef BASE_PROCESSING_HPP
+#define BASE_PROCESSING_HPP
+
+#include <cstddef>
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   Base class exposing pre-processing API.
+     *          Use cases should provide their own PreProcessing class that inherits from this one.
+     *          All steps required to take raw input data and populate tensors ready for inference
+     *          should be handled.
+     */
+    class BasePreProcess {
+
+    public:
+        virtual ~BasePreProcess() = default;
+
+        /**
+         * @brief       Should perform pre-processing of 'raw' input data and load it into
+         *              TFLite Micro input tensors ready for inference
+         * @param[in]   input      Pointer to the data that pre-processing will work on.
+         * @param[in]   inputSize  Size of the input data.
+         * @return      true if successful, false otherwise.
+         **/
+        virtual bool DoPreProcess(const void* input, size_t inputSize) = 0;
+    };
+
+    /**
+     * @brief   Base class exposing post-processing API.
+     *          Use cases should provide their own PostProcessing class that inherits from this one.
+     *          All steps required to take inference output and populate results vectors should be handled.
+     */
+    class BasePostProcess {
+
+    public:
+        virtual ~BasePostProcess() = default;
+
+        /**
+         * @brief       Should perform post-processing of the result of inference then populate
+         *              populate result data for any later use.
+         * @return      true if successful, false otherwise.
+         **/
+        virtual bool DoPostProcess() = 0;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* BASE_PROCESSING_HPP */
\ No newline at end of file
diff --git a/source/application/api/common/include/ClassificationResult.hpp b/source/application/api/common/include/ClassificationResult.hpp
new file mode 100644
index 0000000..eae28e4
--- /dev/null
+++ b/source/application/api/common/include/ClassificationResult.hpp
@@ -0,0 +1,41 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef CLASSIFICATION_RESULT_HPP
+#define CLASSIFICATION_RESULT_HPP
+
+#include <string>
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   Class representing a single classification result.
+     */
+    class ClassificationResult {
+    public:
+        double          m_normalisedVal = 0.0;
+        std::string     m_label;
+        uint32_t        m_labelIdx = 0;
+
+        ClassificationResult() = default;
+        ~ClassificationResult() = default;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* CLASSIFICATION_RESULT_HPP */
\ No newline at end of file
diff --git a/source/application/api/common/include/Classifier.hpp b/source/application/api/common/include/Classifier.hpp
new file mode 100644
index 0000000..d641c22
--- /dev/null
+++ b/source/application/api/common/include/Classifier.hpp
@@ -0,0 +1,89 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef CLASSIFIER_HPP
+#define CLASSIFIER_HPP
+
+#include "ClassificationResult.hpp"
+#include "TensorFlowLiteMicro.hpp"
+
+#include <vector>
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   Classifier - a helper class to get certain number of top
+     *          results from the output vector from a classification NN.
+     **/
+    class Classifier{
+    public:
+        /** @brief Constructor. */
+        Classifier() = default;
+
+        /**
+         * @brief       Gets the top N classification results from the
+         *              output vector.
+         * @param[in]   outputTensor   Inference output tensor from an NN model.
+         * @param[out]  vecResults     A vector of classification results.
+         *                             populated by this function.
+         * @param[in]   labels         Labels vector to match classified classes.
+         * @param[in]   topNCount      Number of top classifications to pick. Default is 1.
+         * @param[in]   useSoftmax     Whether Softmax normalisation should be applied to output. Default is false.
+         * @return      true if successful, false otherwise.
+         **/
+
+        virtual bool GetClassificationResults(
+            TfLiteTensor* outputTensor,
+            std::vector<ClassificationResult>& vecResults,
+            const std::vector <std::string>& labels, uint32_t topNCount,
+            bool use_softmax);
+
+        /**
+        * @brief       Populate the elements of the Classification Result object.
+        * @param[in]   topNSet        Ordered set of top 5 output class scores and labels.
+        * @param[out]  vecResults     A vector of classification results.
+        *                             populated by this function.
+        * @param[in]   labels         Labels vector to match classified classes.
+        **/
+
+        void SetVectorResults(
+            std::set<std::pair<float, uint32_t>>& topNSet,
+            std::vector<ClassificationResult>& vecResults,
+            const std::vector <std::string>& labels);
+
+    private:
+        /**
+         * @brief       Utility function that gets the top N classification results from the
+         *              output vector.
+         * @param[in]   tensor       Inference output tensor from an NN model.
+         * @param[out]  vecResults   A vector of classification results
+         *                           populated by this function.
+         * @param[in]   topNCount    Number of top classifications to pick.
+         * @param[in]   labels       Labels vector to match classified classes.
+         * @return      true if successful, false otherwise.
+         **/
+
+        bool GetTopNResults(const std::vector<float>& tensor,
+                            std::vector<ClassificationResult>& vecResults,
+                            uint32_t topNCount,
+                            const std::vector <std::string>& labels);
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* CLASSIFIER_HPP */
diff --git a/source/application/api/common/include/DataStructures.hpp b/source/application/api/common/include/DataStructures.hpp
new file mode 100644
index 0000000..0616839
--- /dev/null
+++ b/source/application/api/common/include/DataStructures.hpp
@@ -0,0 +1,128 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef DATA_STRUCTURES_HPP
+#define DATA_STRUCTURES_HPP
+
+#include <iterator>
+
+namespace arm {
+namespace app {
+
+    /**
+     * Class Array2d is a data structure that represents a two dimensional array.
+     * The data is allocated in contiguous memory, arranged row-wise
+     * and individual elements can be accessed with the () operator.
+     * For example a two dimensional array D of size (M, N) can be accessed:
+     *
+     *               _|<------------- col size = N  -------->|
+     *               |  D(r=0, c=0) D(r=0, c=1)... D(r=0, c=N)
+     *               |  D(r=1, c=0) D(r=1, c=1)... D(r=1, c=N)
+     *               |  ...
+     *    row size = M  ...
+     *               |  ...
+     *               _  D(r=M, c=0) D(r=M, c=1)... D(r=M, c=N)
+     *
+     */
+    template<typename T>
+    class Array2d {
+    public:
+        /**
+         * @brief     Creates the array2d with the given sizes.
+         * @param[in] rows   Number of rows.
+         * @param[in] cols   Number of columns.
+         */
+        Array2d(unsigned rows, unsigned cols): m_rows(rows), m_cols(cols)
+        {
+            if (rows == 0 || cols == 0) {
+                printf("Array2d constructor has 0 size.\n");
+                m_data = nullptr;
+                return;
+            }
+            m_data = new T[rows * cols];
+        }
+
+        ~Array2d()
+        {
+            delete[] m_data;
+        }
+
+        T& operator() (unsigned int row, unsigned int col)
+        {
+#if defined(DEBUG)
+            if (row >= m_rows || col >= m_cols ||  m_data == nullptr) {
+                printf_err("Array2d subscript out of bounds.\n");
+            }
+#endif /* defined(DEBUG) */
+            return m_data[m_cols * row + col];
+        }
+
+        T operator() (unsigned int row, unsigned int col) const
+        {
+#if defined(DEBUG)
+            if (row >= m_rows || col >= m_cols ||  m_data == nullptr) {
+                printf_err("const Array2d subscript out of bounds.\n");
+            }
+#endif /* defined(DEBUG) */
+            return m_data[m_cols * row + col];
+        }
+
+        /**
+         * @brief  Gets rows number of the current array2d.
+         * @return Number of rows.
+         */
+        size_t size(size_t dim)
+        {
+            switch (dim)
+            {
+                case 0:
+                    return m_rows;
+                case 1:
+                    return m_cols;
+                default:
+                    return 0;
+            }
+        }
+
+        /**
+         * @brief Gets the array2d total size.
+         */
+        size_t totalSize()
+        {
+            return m_rows * m_cols;
+        }
+
+        /**
+         * array2d iterator.
+         */
+        using iterator=T*;
+        using const_iterator=T const*;
+
+        iterator begin() { return m_data; }
+        iterator end() { return m_data + totalSize(); }
+        const_iterator begin() const { return m_data; }
+        const_iterator end() const { return m_data + totalSize(); };
+
+    private:
+        size_t m_rows;
+        size_t m_cols;
+        T* m_data;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* DATA_STRUCTURES_HPP */
\ No newline at end of file
diff --git a/source/application/api/common/include/ImageUtils.hpp b/source/application/api/common/include/ImageUtils.hpp
new file mode 100644
index 0000000..a8c7650
--- /dev/null
+++ b/source/application/api/common/include/ImageUtils.hpp
@@ -0,0 +1,116 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef IMAGE_UTILS_HPP
+#define IMAGE_UTILS_HPP
+
+#include <cstddef>
+#include <cstdint>
+#include <forward_list>
+#include <vector>
+
+/* Helper macro to convert RGB888 to RGB565 format. */
+#define RGB888_TO_RGB565(R8,G8,B8)  ((((R8>>3) & 0x1F) << 11) |     \
+                                     (((G8>>2) & 0x3F) << 5)  |     \
+                                     ((B8>>3) & 0x1F))
+
+constexpr uint16_t COLOR_BLACK  = 0;
+constexpr uint16_t COLOR_GREEN  = RGB888_TO_RGB565(  0, 255,  0); // 2016;
+constexpr uint16_t COLOR_YELLOW = RGB888_TO_RGB565(255, 255,  0); // 65504;
+
+
+namespace arm {
+namespace app {
+namespace image {
+
+    /**
+     * Contains the x,y co-ordinates of a box centre along with the box width and height.
+     */
+    struct Box {
+        float x;
+        float y;
+        float w;
+        float h;
+    };
+
+    struct Detection {
+        Box bbox;
+        std::vector<float> prob;
+        float objectness;
+    };
+
+    /**
+     * @brief       Calculate the 1D overlap.
+     * @param[in]   x1Center   First center point.
+     * @param[in]   width1     First width.
+     * @param[in]   x2Center   Second center point.
+     * @param[in]   width2     Second width.
+     * @return      The overlap between the two lines.
+     **/
+    float Calculate1DOverlap(float x1Center, float width1, float x2Center, float width2);
+
+    /**
+     * @brief       Calculate the intersection between the two given boxes.
+     * @param[in]   box1   First box.
+     * @param[in]   box2   Second box.
+     * @return      The intersection value.
+     **/
+    float CalculateBoxIntersect(Box& box1, Box& box2);
+
+    /**
+     * @brief       Calculate the union between the two given boxes.
+     * @param[in]   box1   First box.
+     * @param[in]   box2   Second box.
+     * @return      The two given boxes union value.
+     **/
+    float CalculateBoxUnion(Box& box1, Box& box2);
+
+    /**
+     * @brief       Calculate the intersection over union between the two given boxes.
+     * @param[in]   box1   First box.
+     * @param[in]   box2   Second box.
+     * @return      The intersection over union value.
+     **/
+    float CalculateBoxIOU(Box& box1, Box& box2);
+
+    /**
+     * @brief       Calculate the Non-Maxima suppression on the given detection boxes.
+     * @param[in]   detections    List of Detection boxes.
+     * @param[in]   classes       Number of classes.
+     * @param[in]   iouThreshold  Intersection over union threshold.
+     **/
+    void CalculateNMS(std::forward_list<Detection>& detections, int classes, float iouThreshold);
+
+    /**
+     * @brief           Helper function to convert a UINT8 image to INT8 format.
+     * @param[in,out]   data            Pointer to the data start.
+     * @param[in]       kMaxImageSize   Total number of pixels in the image.
+     **/
+    void ConvertImgToInt8(void* data, size_t kMaxImageSize);
+
+    /**
+     * @brief       Converts RGB image to grayscale.
+     * @param[in]   srcPtr   Pointer to RGB source image.
+     * @param[out]  dstPtr   Pointer to grayscale destination image.
+     * @param[in]   imgSz    Destination image size.
+     **/
+    void RgbToGrayscale(const uint8_t* srcPtr, uint8_t* dstPtr, size_t dstImgSz);
+
+} /* namespace image */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* IMAGE_UTILS_HPP */
\ No newline at end of file
diff --git a/source/application/api/common/include/Mfcc.hpp b/source/application/api/common/include/Mfcc.hpp
new file mode 100644
index 0000000..86330ca
--- /dev/null
+++ b/source/application/api/common/include/Mfcc.hpp
@@ -0,0 +1,255 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MFCC_HPP
+#define MFCC_HPP
+
+#include "PlatformMath.hpp"
+
+#include <vector>
+#include <cstdint>
+#include <cmath>
+#include <limits>
+#include <string>
+
+namespace arm {
+namespace app {
+namespace audio {
+
+    /* MFCC's consolidated parameters. */
+    class MfccParams {
+    public:
+        float       m_samplingFreq;
+        uint32_t    m_numFbankBins;
+        float       m_melLoFreq;
+        float       m_melHiFreq;
+        uint32_t    m_numMfccFeatures;
+        uint32_t    m_frameLen;
+        uint32_t    m_frameLenPadded;
+        bool        m_useHtkMethod;
+
+        /** @brief  Constructor */
+        MfccParams(float samplingFreq, uint32_t numFbankBins,
+                   float melLoFreq, float melHiFreq,
+                   uint32_t numMfccFeats, uint32_t frameLen,
+                   bool useHtkMethod);
+
+        MfccParams()  = delete;
+
+        ~MfccParams() = default;
+
+        /** @brief  Log parameters */
+        void Log() const;
+    };
+
+    /**
+     * @brief   Class for MFCC feature extraction.
+     *          Based on https://github.com/ARM-software/ML-KWS-for-MCU/blob/master/Deployment/Source/MFCC/mfcc.cpp
+     *          This class is designed to be generic and self-sufficient but
+     *          certain calculation routines can be overridden to accommodate
+     *          use-case specific requirements.
+     */
+    class MFCC {
+    public:
+        /**
+         * @brief       Constructor
+         * @param[in]   params   MFCC parameters
+        */
+        explicit MFCC(const MfccParams& params);
+
+        MFCC() = delete;
+
+        ~MFCC() = default;
+
+        /**
+        * @brief        Extract MFCC  features for one single small frame of
+        *               audio data e.g. 640 samples.
+        * @param[in]    audioData   Vector of audio samples to calculate
+        *                           features for.
+        * @return       Vector of extracted MFCC features.
+        **/
+        std::vector<float> MfccCompute(const std::vector<int16_t>& audioData);
+
+        /** @brief  Initialise. */
+        void Init();
+
+       /**
+        * @brief        Extract MFCC features and quantise for one single small
+        *               frame of audio data e.g. 640 samples.
+        * @param[in]    audioData     Vector of audio samples to calculate
+        *                             features for.
+        * @param[in]    quantScale    Quantisation scale.
+        * @param[in]    quantOffset   Quantisation offset.
+        * @return       Vector of extracted quantised MFCC features.
+        **/
+        template<typename T>
+        std::vector<T> MfccComputeQuant(const std::vector<int16_t>& audioData,
+                                        const float quantScale,
+                                        const int quantOffset)
+        {
+            this->MfccComputePreFeature(audioData);
+            float minVal = std::numeric_limits<T>::min();
+            float maxVal = std::numeric_limits<T>::max();
+
+            std::vector<T> mfccOut(this->m_params.m_numMfccFeatures);
+            const size_t numFbankBins = this->m_params.m_numFbankBins;
+
+            /* Take DCT. Uses matrix mul. */
+            for (size_t i = 0, j = 0; i < mfccOut.size(); ++i, j += numFbankBins) {
+                float sum = 0;
+                for (size_t k = 0; k < numFbankBins; ++k) {
+                    sum += this->m_dctMatrix[j + k] * this->m_melEnergies[k];
+                }
+                /* Quantize to T. */
+                sum = std::round((sum / quantScale) + quantOffset);
+                mfccOut[i] = static_cast<T>(std::min<float>(std::max<float>(sum, minVal), maxVal));
+            }
+
+            return mfccOut;
+        }
+
+        /* Constants */
+        static constexpr float ms_logStep = /*logf(6.4)*/ 1.8562979903656 / 27.0;
+        static constexpr float ms_freqStep = 200.0 / 3;
+        static constexpr float ms_minLogHz = 1000.0;
+        static constexpr float ms_minLogMel = ms_minLogHz / ms_freqStep;
+
+    protected:
+        /**
+         * @brief       Project input frequency to Mel Scale.
+         * @param[in]   freq           Input frequency in floating point.
+         * @param[in]   useHTKMethod   bool to signal if HTK method is to be
+         *                             used for calculation.
+         * @return      Mel transformed frequency in floating point.
+         **/
+        static float MelScale(float freq,
+                              bool  useHTKMethod = true);
+
+        /**
+         * @brief       Inverse Mel transform - convert MEL warped frequency
+         *              back to normal frequency.
+         * @param[in]   melFreq        Mel frequency in floating point.
+         * @param[in]   useHTKMethod   bool to signal if HTK method is to be
+         *                             used for calculation.
+         * @return      Real world frequency in floating point.
+         **/
+        static float InverseMelScale(float melFreq,
+                                     bool  useHTKMethod = true);
+
+        /**
+         * @brief       Populates MEL energies after applying the MEL filter
+         *              bank weights and adding them up to be placed into
+         *              bins, according to the filter bank's first and last
+         *              indices (pre-computed for each filter bank element
+         *              by CreateMelFilterBank function).
+         * @param[in]   fftVec                  Vector populated with FFT magnitudes.
+         * @param[in]   melFilterBank           2D Vector with filter bank weights.
+         * @param[in]   filterBankFilterFirst   Vector containing the first indices of filter bank
+         *                                      to be used for each bin.
+         * @param[in]   filterBankFilterLast    Vector containing the last indices of filter bank
+         *                                      to be used for each bin.
+         * @param[out]  melEnergies             Pre-allocated vector of MEL energies to be
+         *                                      populated.
+         * @return      true if successful, false otherwise.
+         */
+        virtual bool ApplyMelFilterBank(
+            std::vector<float>&                 fftVec,
+            std::vector<std::vector<float>>&    melFilterBank,
+            std::vector<uint32_t>&               filterBankFilterFirst,
+            std::vector<uint32_t>&               filterBankFilterLast,
+            std::vector<float>&                 melEnergies);
+
+        /**
+         * @brief           Converts the Mel energies for logarithmic scale.
+         * @param[in,out]   melEnergies   1D vector of Mel energies.
+         **/
+        virtual void ConvertToLogarithmicScale(std::vector<float>& melEnergies);
+
+        /**
+         * @brief       Create a matrix used to calculate Discrete Cosine
+         *              Transform.
+         * @param[in]   inputLength        Input length of the buffer on which
+         *                                 DCT will be performed.
+         * @param[in]   coefficientCount   Total coefficients per input length.
+         * @return      1D vector with inputLength x coefficientCount elements
+         *              populated with DCT coefficients.
+         */
+        virtual std::vector<float> CreateDCTMatrix(
+                                    int32_t inputLength,
+                                    int32_t coefficientCount);
+
+        /**
+         * @brief       Given the low and high Mel values, get the normaliser
+         *              for weights to be applied when populating the filter
+         *              bank.
+         * @param[in]   leftMel        Low Mel frequency value.
+         * @param[in]   rightMel       High Mel frequency value.
+         * @param[in]   useHTKMethod   bool to signal if HTK method is to be
+         *                             used for calculation.
+         * @return      Value to use for normalizing.
+         */
+        virtual float GetMelFilterBankNormaliser(
+                        const float&   leftMel,
+                        const float&   rightMel,
+                        bool     useHTKMethod);
+
+    private:
+        MfccParams                      m_params;
+        std::vector<float>              m_frame;
+        std::vector<float>              m_buffer;
+        std::vector<float>              m_melEnergies;
+        std::vector<float>              m_windowFunc;
+        std::vector<std::vector<float>> m_melFilterBank;
+        std::vector<float>              m_dctMatrix;
+        std::vector<uint32_t>           m_filterBankFilterFirst;
+        std::vector<uint32_t>           m_filterBankFilterLast;
+        bool                            m_filterBankInitialised;
+        arm::app::math::FftInstance     m_fftInstance;
+
+        /**
+         * @brief       Initialises the filter banks and the DCT matrix. **/
+        void InitMelFilterBank();
+
+        /**
+         * @brief       Signals whether the instance of MFCC has had its
+         *              required buffers initialised.
+         * @return      true if initialised, false otherwise.
+         **/
+        bool IsMelFilterBankInited() const;
+
+        /**
+         * @brief       Create mel filter banks for MFCC calculation.
+         * @return      2D vector of floats.
+         **/
+        std::vector<std::vector<float>> CreateMelFilterBank();
+
+        /**
+         * @brief       Computes and populates internal memeber buffers used
+         *              in MFCC feature calculation
+         * @param[in]   audioData   1D vector of 16-bit audio data.
+         */
+        void MfccComputePreFeature(const std::vector<int16_t>& audioData);
+
+        /** @brief       Computes the magnitude from an interleaved complex array. */
+        void ConvertToPowerSpectrum();
+
+    };
+
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* MFCC_HPP */
\ No newline at end of file
diff --git a/source/application/api/common/include/Model.hpp b/source/application/api/common/include/Model.hpp
new file mode 100644
index 0000000..df1b259
--- /dev/null
+++ b/source/application/api/common/include/Model.hpp
@@ -0,0 +1,152 @@
+/*
+ * Copyright (c) 2021-2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MODEL_HPP
+#define MODEL_HPP
+
+#include "TensorFlowLiteMicro.hpp"
+
+#include <cstdint>
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   NN model class wrapping the underlying TensorFlow-Lite-Micro API.
+     */
+    class Model {
+    public:
+        /** @brief Constructor. */
+        Model();
+
+        /** @brief Destructor. */
+        ~Model();
+
+        /** @brief  Gets the pointer to the model's input tensor at given input index. */
+        TfLiteTensor* GetInputTensor(size_t index) const;
+
+        /** @brief  Gets the pointer to the model's output tensor at given output index. */
+        TfLiteTensor* GetOutputTensor(size_t index) const;
+
+        /** @brief  Gets the model's data type. */
+        TfLiteType GetType() const;
+
+        /** @brief  Gets the pointer to the model's input shape. */
+        TfLiteIntArray* GetInputShape(size_t index) const;
+
+        /** @brief  Gets the pointer to the model's output shape at given output index. */
+        TfLiteIntArray* GetOutputShape(size_t index) const;
+
+        /** @brief  Gets the number of input tensors the model has. */
+        size_t GetNumInputs() const;
+
+        /** @brief  Gets the number of output tensors the model has. */
+        size_t GetNumOutputs() const;
+
+        /** @brief  Logs the tensor information to stdout. */
+        void LogTensorInfo(TfLiteTensor* tensor);
+
+        /** @brief  Logs the interpreter information to stdout. */
+        void LogInterpreterInfo();
+
+        /** @brief      Initialise the model class object.
+         *  @param[in]  tensorArenaAddress  Pointer to the tensor arena buffer.
+         *  @param[in]  tensorArenaAddress  Size of the tensor arena buffer in bytes.
+         *  @param[in]  nnModelAddr         Pointer to the model.
+         *  @param[in]  nnModelSize         Size of the model in bytes, if known.
+         *  @param[in]  allocator   Optional: a pre-initialised micro allocator pointer,
+         *                          if available. If supplied, this allocator will be used
+         *                          to create the interpreter instance.
+         *  @return     true if initialisation succeeds, false otherwise.
+        **/
+        bool Init(uint8_t* tensorArenaAddr,
+                  uint32_t tensorArenaSize,
+                  uint8_t* nnModelAddr,
+                  uint32_t nnModelSize,
+                  tflite::MicroAllocator* allocator = nullptr);
+
+        /**
+         * @brief       Gets the allocator pointer for this instance.
+         * @return      Pointer to a tflite::MicroAllocator object, if
+         *              available; nullptr otherwise.
+         **/
+        tflite::MicroAllocator* GetAllocator();
+
+        /** @brief  Checks if this object has been initialised. */
+        bool IsInited() const;
+
+        /** @brief  Checks if the model uses signed data. */
+        bool IsDataSigned() const;
+
+        /** @brief Checks if the model uses Ethos-U operator */
+        bool ContainsEthosUOperator() const;
+
+        /** @brief  Runs the inference (invokes the interpreter). */
+        virtual bool RunInference();
+
+        /** @brief   Model information handler common to all models.
+         *  @return  true or false based on execution success.
+         **/
+        bool ShowModelInfoHandler();
+
+        /** @brief   Gets a pointer to the tensor arena. */
+        uint8_t* GetTensorArena();
+
+    protected:
+        /** @brief      Gets the pointer to the NN model data array.
+         *  @return     Pointer of uint8_t type.
+         **/
+        const uint8_t* ModelPointer();
+
+        /** @brief      Gets the model size.
+         *  @return     size_t, size in bytes.
+         **/
+        uint32_t ModelSize();
+
+        /**
+         * @brief       Gets the op resolver for the model instance.
+         * @return      const reference to a tflite::MicroOpResolver object.
+         **/
+        virtual const tflite::MicroOpResolver& GetOpResolver() = 0;
+
+        /**
+         * @brief       Add all the operators required for the given model.
+         *              Implementation of this should come from the use case.
+         * @return      true is ops are successfully added, false otherwise.
+         **/
+        virtual bool EnlistOperations() = 0;
+
+        /** @brief   Gets the total size of tensor arena available for use. */
+        size_t GetActivationBufferSize();
+
+    private:
+        tflite::ErrorReporter*          m_pErrorReporter     = nullptr;     /* Pointer to the error reporter. */
+        const tflite::Model*            m_pModel             = nullptr;     /* Tflite model pointer. */
+        tflite::MicroInterpreter*       m_pInterpreter       = nullptr;     /* Tflite interpreter. */
+        tflite::MicroAllocator*         m_pAllocator         = nullptr;     /* Tflite micro allocator. */
+        bool                            m_inited             = false;       /* Indicates whether this object has been initialised. */
+        uint8_t*                        m_modelAddr          = nullptr;     /* Model address */
+        uint32_t                        m_modelSize          = 0;           /* Model size */
+
+        std::vector<TfLiteTensor*>      m_input              = {};           /* Model's input tensor pointers. */
+        std::vector<TfLiteTensor*>      m_output             = {};           /* Model's output tensor pointers. */
+        TfLiteType                      m_type               = kTfLiteNoType;/* Model's data type. */
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* MODEL_HPP */
diff --git a/source/application/api/common/include/TensorFlowLiteMicro.hpp b/source/application/api/common/include/TensorFlowLiteMicro.hpp
new file mode 100644
index 0000000..f6639fd
--- /dev/null
+++ b/source/application/api/common/include/TensorFlowLiteMicro.hpp
@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef TENSORFLOW_LITE_MICRO_LOCAL_HPP
+#define TENSORFLOW_LITE_MICRO_LOCAL_HPP
+
+/* We include all our TensorFlow Lite Micro headers here */
+
+/**
+ * TensorFlow Lite Micro sources can generate a lot of warnings from the usage
+ * of a single macro (TF_LITE_REMOVE_VIRTUAL_DELETE). Suppress the known ones
+ * here to prevent them from masking warnings that might be generated by our
+ * application sources.
+ */
+#if defined(__ARMCC_VERSION) && (__ARMCC_VERSION >= 6010050)
+    #pragma clang diagnostic push
+    #pragma clang diagnostic ignored "-Wunused-parameter"
+    #include "tensorflow/lite/micro/micro_mutable_op_resolver.h"
+    #include "tensorflow/lite/micro/micro_interpreter.h"
+    #include "tensorflow/lite/micro/micro_error_reporter.h"
+    #include "tensorflow/lite/micro/all_ops_resolver.h"
+    #pragma clang diagnostic pop
+#elif defined(__GNUC__)
+    #pragma GCC diagnostic push
+    #pragma GCC diagnostic ignored "-Wunused-parameter"
+    #include "tensorflow/lite/micro/micro_mutable_op_resolver.h"
+    #include "tensorflow/lite/micro/micro_interpreter.h"
+    #include "tensorflow/lite/micro/micro_error_reporter.h"
+    #include "tensorflow/lite/micro/all_ops_resolver.h"
+    #pragma GCC diagnostic pop
+#else
+    #include "tensorflow/lite/micro/micro_mutable_op_resolver.h"
+    #include "tensorflow/lite/micro/micro_interpreter.h"
+    #include "tensorflow/lite/micro/micro_error_reporter.h"
+    #include "tensorflow/lite/micro/all_ops_resolver.h"
+#endif
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/micro/kernels/micro_ops.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/schema/schema_utils.h"
+
+#if defined (TESTS)
+    #include "tensorflow/lite/micro/test_helpers.h"
+#endif /* defined (TESTS) */
+
+namespace arm {
+namespace app {
+
+    /** Struct for quantization parameters. */
+    struct QuantParams {
+        float   scale   = 1.0;
+        int     offset  = 0;
+    };
+
+    /**
+     * @brief       Gets the quantization parameters from a tensor
+     * @param[in]   tensor  pointer to the tensor.
+     * @return      QuantParams object.
+     */
+    QuantParams GetTensorQuantParams(TfLiteTensor* tensor);
+
+    /**
+     * @brief   String logging functionality expected to be defined
+     *          by TensorFlow Lite Micro's error reporter.
+     * @param[in]   s   Pointer to the string.
+     */
+    extern "C" void DebugLog(const char* s);
+
+} /* namespace app */
+} /* namespace arm */
+
+/**
+ * @brief Prints the tensor flow version in use to stdout.
+ */
+void PrintTensorFlowVersion();
+
+#endif /* TENSORFLOW_LITE_MICRO_LOCAL_HPP */
diff --git a/source/application/api/common/source/Classifier.cc b/source/application/api/common/source/Classifier.cc
new file mode 100644
index 0000000..6fabebe
--- /dev/null
+++ b/source/application/api/common/source/Classifier.cc
@@ -0,0 +1,169 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "Classifier.hpp"
+
+#include "TensorFlowLiteMicro.hpp"
+#include "PlatformMath.hpp"
+#include "log_macros.h"
+
+#include <vector>
+#include <string>
+#include <set>
+#include <cstdint>
+#include <cinttypes>
+
+
+namespace arm {
+namespace app {
+
+    void Classifier::SetVectorResults(std::set<std::pair<float, uint32_t>>& topNSet,
+                          std::vector<ClassificationResult>& vecResults,
+                          const std::vector <std::string>& labels)
+    {
+
+        /* Reset the iterator to the largest element - use reverse iterator. */
+
+        auto topNIter = topNSet.rbegin();
+        for (size_t i = 0; i < vecResults.size() && topNIter != topNSet.rend(); ++i, ++topNIter) {
+            vecResults[i].m_normalisedVal = topNIter->first;
+            vecResults[i].m_label = labels[topNIter->second];
+            vecResults[i].m_labelIdx = topNIter->second;
+        }
+    }
+
+    bool Classifier::GetTopNResults(const std::vector<float>& tensor,
+                                    std::vector<ClassificationResult>& vecResults,
+                                    uint32_t topNCount,
+                                    const std::vector <std::string>& labels)
+    {
+
+        std::set<std::pair<float , uint32_t>> sortedSet;
+
+        /* NOTE: inputVec's size verification against labels should be
+         *       checked by the calling/public function. */
+
+        /* Set initial elements. */
+        for (uint32_t i = 0; i < topNCount; ++i) {
+            sortedSet.insert({tensor[i], i});
+        }
+
+        /* Initialise iterator. */
+        auto setFwdIter = sortedSet.begin();
+
+        /* Scan through the rest of elements with compare operations. */
+        for (uint32_t i = topNCount; i < labels.size(); ++i) {
+            if (setFwdIter->first < tensor[i]) {
+                sortedSet.erase(*setFwdIter);
+                sortedSet.insert({tensor[i], i});
+                setFwdIter = sortedSet.begin();
+            }
+        }
+
+        /* Final results' container. */
+        vecResults = std::vector<ClassificationResult>(topNCount);
+        SetVectorResults(sortedSet, vecResults, labels);
+
+        return true;
+    }
+
+    bool  Classifier::GetClassificationResults(
+        TfLiteTensor* outputTensor,
+        std::vector<ClassificationResult>& vecResults,
+        const std::vector <std::string>& labels,
+        uint32_t topNCount,
+        bool useSoftmax)
+    {
+        if (outputTensor == nullptr) {
+            printf_err("Output vector is null pointer.\n");
+            return false;
+        }
+
+        uint32_t totalOutputSize = 1;
+        for (int inputDim = 0; inputDim < outputTensor->dims->size; inputDim++) {
+            totalOutputSize *= outputTensor->dims->data[inputDim];
+        }
+
+        /* Sanity checks. */
+        if (totalOutputSize < topNCount) {
+            printf_err("Output vector is smaller than %" PRIu32 "\n", topNCount);
+            return false;
+        } else if (totalOutputSize != labels.size()) {
+            printf_err("Output size doesn't match the labels' size\n");
+            return false;
+        } else if (topNCount == 0) {
+            printf_err("Top N results cannot be zero\n");
+            return false;
+        }
+
+        bool resultState;
+        vecResults.clear();
+
+        /* De-Quantize Output Tensor */
+        QuantParams quantParams = GetTensorQuantParams(outputTensor);
+
+        /* Floating point tensor data to be populated
+         * NOTE: The assumption here is that the output tensor size isn't too
+         * big and therefore, there's neglibible impact on heap usage. */
+        std::vector<float> tensorData(totalOutputSize);
+
+        /* Populate the floating point buffer */
+        switch (outputTensor->type) {
+            case kTfLiteUInt8: {
+                uint8_t *tensor_buffer = tflite::GetTensorData<uint8_t>(outputTensor);
+                for (size_t i = 0; i < totalOutputSize; ++i) {
+                    tensorData[i] = quantParams.scale *
+                        (static_cast<float>(tensor_buffer[i]) - quantParams.offset);
+                }
+                break;
+            }
+            case kTfLiteInt8: {
+                int8_t *tensor_buffer = tflite::GetTensorData<int8_t>(outputTensor);
+                for (size_t i = 0; i < totalOutputSize; ++i) {
+                    tensorData[i] = quantParams.scale *
+                        (static_cast<float>(tensor_buffer[i]) - quantParams.offset);
+                }
+                break;
+            }
+            case kTfLiteFloat32: {
+                float *tensor_buffer = tflite::GetTensorData<float>(outputTensor);
+                for (size_t i = 0; i < totalOutputSize; ++i) {
+                    tensorData[i] = tensor_buffer[i];
+                }
+                break;
+            }
+            default:
+                printf_err("Tensor type %s not supported by classifier\n",
+                    TfLiteTypeGetName(outputTensor->type));
+                return false;
+        }
+
+        if (useSoftmax) {
+            math::MathUtils::SoftmaxF32(tensorData);
+        }
+
+        /* Get the top N results. */
+        resultState = GetTopNResults(tensorData, vecResults, topNCount, labels);
+
+        if (!resultState) {
+            printf_err("Failed to get top N results set\n");
+            return false;
+        }
+
+        return true;
+    }
+} /* namespace app */
+} /* namespace arm */
\ No newline at end of file
diff --git a/source/application/api/common/source/ImageUtils.cc b/source/application/api/common/source/ImageUtils.cc
new file mode 100644
index 0000000..31b9493
--- /dev/null
+++ b/source/application/api/common/source/ImageUtils.cc
@@ -0,0 +1,126 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "ImageUtils.hpp"
+
+#include <limits>
+
+namespace arm {
+namespace app {
+namespace image {
+
+    float Calculate1DOverlap(float x1Center, float width1, float x2Center, float width2)
+    {
+        float left_1 = x1Center - width1/2;
+        float left_2 = x2Center - width2/2;
+        float leftest = left_1 > left_2 ? left_1 : left_2;
+
+        float right_1 = x1Center + width1/2;
+        float right_2 = x2Center + width2/2;
+        float rightest = right_1 < right_2 ? right_1 : right_2;
+
+        return rightest - leftest;
+    }
+
+    float CalculateBoxIntersect(Box& box1, Box& box2)
+    {
+        float width = Calculate1DOverlap(box1.x, box1.w, box2.x, box2.w);
+        if (width < 0) {
+            return 0;
+        }
+        float height = Calculate1DOverlap(box1.y, box1.h, box2.y, box2.h);
+        if (height < 0) {
+            return 0;
+        }
+
+        float total_area = width*height;
+        return total_area;
+    }
+
+    float CalculateBoxUnion(Box& box1, Box& box2)
+    {
+        float boxes_intersection = CalculateBoxIntersect(box1, box2);
+        float boxes_union = box1.w * box1.h + box2.w * box2.h - boxes_intersection;
+        return boxes_union;
+    }
+
+    float CalculateBoxIOU(Box& box1, Box& box2)
+    {
+        float boxes_intersection = CalculateBoxIntersect(box1, box2);
+        if (boxes_intersection == 0) {
+            return 0;
+        }
+
+        float boxes_union = CalculateBoxUnion(box1, box2);
+        if (boxes_union == 0) {
+            return 0;
+        }
+
+        return boxes_intersection / boxes_union;
+    }
+
+    void CalculateNMS(std::forward_list<Detection>& detections, int classes, float iouThreshold)
+    {
+        int idxClass{0};
+        auto CompareProbs = [idxClass](Detection& prob1, Detection& prob2) {
+            return prob1.prob[idxClass] > prob2.prob[idxClass];
+        };
+
+        for (idxClass = 0; idxClass < classes; ++idxClass) {
+            detections.sort(CompareProbs);
+
+            for (auto it=detections.begin(); it != detections.end(); ++it) {
+                if (it->prob[idxClass] == 0) continue;
+                for (auto itc=std::next(it, 1); itc != detections.end(); ++itc) {
+                    if (itc->prob[idxClass] == 0) {
+                        continue;
+                    }
+                    if (CalculateBoxIOU(it->bbox, itc->bbox) > iouThreshold) {
+                        itc->prob[idxClass] = 0;
+                    }
+                }
+            }
+        }
+    }
+
+    void ConvertImgToInt8(void* data, const size_t kMaxImageSize)
+    {
+        auto* tmp_req_data = static_cast<uint8_t*>(data);
+        auto* tmp_signed_req_data = static_cast<int8_t*>(data);
+
+        for (size_t i = 0; i < kMaxImageSize; i++) {
+            tmp_signed_req_data[i] = (int8_t) (
+                    (int32_t) (tmp_req_data[i]) - 128);
+        }
+    }
+
+    void RgbToGrayscale(const uint8_t* srcPtr, uint8_t* dstPtr, const size_t dstImgSz)
+    {
+        const float R = 0.299;
+        const float G = 0.587;
+        const float B = 0.114;
+        for (size_t i = 0; i < dstImgSz; ++i, srcPtr += 3) {
+            uint32_t  int_gray = R * (*srcPtr) +
+                                 G * (*(srcPtr + 1)) +
+                                 B * (*(srcPtr + 2));
+            *dstPtr++ = int_gray <= std::numeric_limits<uint8_t>::max() ?
+                        int_gray : std::numeric_limits<uint8_t>::max();
+        }
+    }
+
+} /* namespace image */
+} /* namespace app */
+} /* namespace arm */
\ No newline at end of file
diff --git a/source/application/api/common/source/Mfcc.cc b/source/application/api/common/source/Mfcc.cc
new file mode 100644
index 0000000..3bf5eb3
--- /dev/null
+++ b/source/application/api/common/source/Mfcc.cc
@@ -0,0 +1,353 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "Mfcc.hpp"
+#include "PlatformMath.hpp"
+#include "log_macros.h"
+
+#include <cfloat>
+#include <cinttypes>
+
+namespace arm {
+namespace app {
+namespace audio {
+
+    MfccParams::MfccParams(
+                    const float samplingFreq,
+                    const uint32_t numFbankBins,
+                    const float melLoFreq,
+                    const float melHiFreq,
+                    const uint32_t numMfccFeats,
+                    const uint32_t frameLen,
+                    const bool useHtkMethod):
+                        m_samplingFreq(samplingFreq),
+                        m_numFbankBins(numFbankBins),
+                        m_melLoFreq(melLoFreq),
+                        m_melHiFreq(melHiFreq),
+                        m_numMfccFeatures(numMfccFeats),
+                        m_frameLen(frameLen),
+
+                        /* Smallest power of 2 >= frame length. */
+                        m_frameLenPadded(pow(2, ceil((log(frameLen)/log(2))))),
+                        m_useHtkMethod(useHtkMethod)
+    {}
+
+    void MfccParams::Log() const
+    {
+        debug("MFCC parameters:\n");
+        debug("\t Sampling frequency:         %f\n", this->m_samplingFreq);
+        debug("\t Number of filter banks:     %" PRIu32 "\n", this->m_numFbankBins);
+        debug("\t Mel frequency limit (low):  %f\n", this->m_melLoFreq);
+        debug("\t Mel frequency limit (high): %f\n", this->m_melHiFreq);
+        debug("\t Number of MFCC features:    %" PRIu32 "\n", this->m_numMfccFeatures);
+        debug("\t Frame length:               %" PRIu32 "\n", this->m_frameLen);
+        debug("\t Padded frame length:        %" PRIu32 "\n", this->m_frameLenPadded);
+        debug("\t Using HTK for Mel scale:    %s\n", this->m_useHtkMethod ? "yes" : "no");
+    }
+
+    MFCC::MFCC(const MfccParams& params):
+        m_params(params),
+        m_filterBankInitialised(false)
+    {
+        this->m_buffer = std::vector<float>(
+                            this->m_params.m_frameLenPadded, 0.0);
+        this->m_frame = std::vector<float>(
+                            this->m_params.m_frameLenPadded, 0.0);
+        this->m_melEnergies = std::vector<float>(
+                                this->m_params.m_numFbankBins, 0.0);
+
+        this->m_windowFunc = std::vector<float>(this->m_params.m_frameLen);
+        const auto multiplier = static_cast<float>(2 * M_PI / this->m_params.m_frameLen);
+
+        /* Create window function. */
+        for (size_t i = 0; i < this->m_params.m_frameLen; i++) {
+            this->m_windowFunc[i] = (0.5 - (0.5 *
+                math::MathUtils::CosineF32(static_cast<float>(i) * multiplier)));
+        }
+
+        math::MathUtils::FftInitF32(this->m_params.m_frameLenPadded, this->m_fftInstance);
+        this->m_params.Log();
+    }
+
+    void MFCC::Init()
+    {
+        this->InitMelFilterBank();
+    }
+
+    float MFCC::MelScale(const float freq, const bool useHTKMethod)
+    {
+        if (useHTKMethod) {
+            return 1127.0f * logf (1.0f + freq / 700.0f);
+        } else {
+            /* Slaney formula for mel scale. */
+
+            float mel = freq / ms_freqStep;
+
+            if (freq >= ms_minLogHz) {
+                mel = ms_minLogMel + logf(freq / ms_minLogHz) / ms_logStep;
+            }
+            return mel;
+        }
+    }
+
+    float MFCC::InverseMelScale(const float melFreq, const bool useHTKMethod)
+    {
+        if (useHTKMethod) {
+            return 700.0f * (expf (melFreq / 1127.0f) - 1.0f);
+        } else {
+            /* Slaney formula for mel scale. */
+            float freq = ms_freqStep * melFreq;
+
+            if (melFreq >= ms_minLogMel) {
+                freq = ms_minLogHz * expf(ms_logStep * (melFreq - ms_minLogMel));
+            }
+            return freq;
+        }
+    }
+
+
+    bool MFCC::ApplyMelFilterBank(
+            std::vector<float>&                 fftVec,
+            std::vector<std::vector<float>>&    melFilterBank,
+            std::vector<uint32_t>&               filterBankFilterFirst,
+            std::vector<uint32_t>&               filterBankFilterLast,
+            std::vector<float>&                 melEnergies)
+    {
+        const size_t numBanks = melEnergies.size();
+
+        if (numBanks != filterBankFilterFirst.size() ||
+                numBanks != filterBankFilterLast.size()) {
+            printf_err("unexpected filter bank lengths\n");
+            return false;
+        }
+
+        for (size_t bin = 0; bin < numBanks; ++bin) {
+            auto filterBankIter = melFilterBank[bin].begin();
+            auto end = melFilterBank[bin].end();
+            float melEnergy = FLT_MIN;  /* Avoid log of zero at later stages */
+            const uint32_t firstIndex = filterBankFilterFirst[bin];
+            const uint32_t lastIndex = std::min<uint32_t>(filterBankFilterLast[bin], fftVec.size() - 1);
+
+            for (uint32_t i = firstIndex; i <= lastIndex && filterBankIter != end; i++) {
+                float energyRep = math::MathUtils::SqrtF32(fftVec[i]);
+                melEnergy += (*filterBankIter++ * energyRep);
+            }
+
+            melEnergies[bin] = melEnergy;
+        }
+
+        return true;
+    }
+
+    void MFCC::ConvertToLogarithmicScale(std::vector<float>& melEnergies)
+    {
+        for (float& melEnergy : melEnergies) {
+            melEnergy = logf(melEnergy);
+        }
+    }
+
+    void MFCC::ConvertToPowerSpectrum()
+    {
+        const uint32_t halfDim = this->m_buffer.size() / 2;
+
+        /* Handle this special case. */
+        float firstEnergy = this->m_buffer[0] * this->m_buffer[0];
+        float lastEnergy = this->m_buffer[1] * this->m_buffer[1];
+
+        math::MathUtils::ComplexMagnitudeSquaredF32(
+                            this->m_buffer.data(),
+                            this->m_buffer.size(),
+                            this->m_buffer.data(),
+                            this->m_buffer.size()/2);
+
+        this->m_buffer[0] = firstEnergy;
+        this->m_buffer[halfDim] = lastEnergy;
+    }
+
+    std::vector<float> MFCC::CreateDCTMatrix(
+                                const int32_t inputLength,
+                                const int32_t coefficientCount)
+    {
+        std::vector<float> dctMatix(inputLength * coefficientCount);
+
+        const float normalizer = math::MathUtils::SqrtF32(2.0f/inputLength);
+        const float angleIncr = M_PI/inputLength;
+        float angle = 0;
+
+        for (int32_t k = 0, m = 0; k < coefficientCount; k++, m += inputLength) {
+            for (int32_t n = 0; n < inputLength; n++) {
+                dctMatix[m+n] = normalizer *
+                    math::MathUtils::CosineF32((n + 0.5f) * angle);
+            }
+            angle += angleIncr;
+        }
+
+        return dctMatix;
+    }
+
+    float MFCC::GetMelFilterBankNormaliser(
+                    const float&    leftMel,
+                    const float&    rightMel,
+                    const bool      useHTKMethod)
+    {
+        UNUSED(leftMel);
+        UNUSED(rightMel);
+        UNUSED(useHTKMethod);
+
+        /* By default, no normalisation => return 1 */
+        return 1.f;
+    }
+
+    void MFCC::InitMelFilterBank()
+    {
+        if (!this->IsMelFilterBankInited()) {
+            this->m_melFilterBank = this->CreateMelFilterBank();
+            this->m_dctMatrix = this->CreateDCTMatrix(
+                                    this->m_params.m_numFbankBins,
+                                    this->m_params.m_numMfccFeatures);
+            this->m_filterBankInitialised = true;
+        }
+    }
+
+    bool MFCC::IsMelFilterBankInited() const
+    {
+        return this->m_filterBankInitialised;
+    }
+
+    void MFCC::MfccComputePreFeature(const std::vector<int16_t>& audioData)
+    {
+        this->InitMelFilterBank();
+
+        /* TensorFlow way of normalizing .wav data to (-1, 1). */
+        constexpr float normaliser = 1.0/(1u<<15u);
+        for (size_t i = 0; i < this->m_params.m_frameLen; i++) {
+            this->m_frame[i] = static_cast<float>(audioData[i]) * normaliser;
+        }
+
+        /* Apply window function to input frame. */
+        for(size_t i = 0; i < this->m_params.m_frameLen; i++) {
+            this->m_frame[i] *= this->m_windowFunc[i];
+        }
+
+        /* Set remaining frame values to 0. */
+        std::fill(this->m_frame.begin() + this->m_params.m_frameLen,this->m_frame.end(), 0);
+
+        /* Compute FFT. */
+        math::MathUtils::FftF32(this->m_frame, this->m_buffer, this->m_fftInstance);
+
+        /* Convert to power spectrum. */
+        this->ConvertToPowerSpectrum();
+
+        /* Apply mel filterbanks. */
+        if (!this->ApplyMelFilterBank(this->m_buffer,
+                                      this->m_melFilterBank,
+                                      this->m_filterBankFilterFirst,
+                                      this->m_filterBankFilterLast,
+                                      this->m_melEnergies)) {
+            printf_err("Failed to apply MEL filter banks\n");
+        }
+
+        /* Convert to logarithmic scale. */
+        this->ConvertToLogarithmicScale(this->m_melEnergies);
+    }
+
+    std::vector<float> MFCC::MfccCompute(const std::vector<int16_t>& audioData)
+    {
+        this->MfccComputePreFeature(audioData);
+
+        std::vector<float> mfccOut(this->m_params.m_numMfccFeatures);
+
+        float * ptrMel = this->m_melEnergies.data();
+        float * ptrDct = this->m_dctMatrix.data();
+        float * ptrMfcc = mfccOut.data();
+
+        /* Take DCT. Uses matrix mul. */
+        for (size_t i = 0, j = 0; i < mfccOut.size();
+                    ++i, j += this->m_params.m_numFbankBins) {
+            *ptrMfcc++ = math::MathUtils::DotProductF32(
+                                            ptrDct + j,
+                                            ptrMel,
+                                            this->m_params.m_numFbankBins);
+        }
+        return mfccOut;
+    }
+
+    std::vector<std::vector<float>> MFCC::CreateMelFilterBank()
+    {
+        size_t numFftBins = this->m_params.m_frameLenPadded / 2;
+        float fftBinWidth = static_cast<float>(this->m_params.m_samplingFreq) / this->m_params.m_frameLenPadded;
+
+        float melLowFreq = MFCC::MelScale(this->m_params.m_melLoFreq,
+                                          this->m_params.m_useHtkMethod);
+        float melHighFreq = MFCC::MelScale(this->m_params.m_melHiFreq,
+                                           this->m_params.m_useHtkMethod);
+        float melFreqDelta = (melHighFreq - melLowFreq) / (this->m_params.m_numFbankBins + 1);
+
+        std::vector<float> thisBin = std::vector<float>(numFftBins);
+        std::vector<std::vector<float>> melFilterBank(
+                                            this->m_params.m_numFbankBins);
+        this->m_filterBankFilterFirst =
+                        std::vector<uint32_t>(this->m_params.m_numFbankBins);
+        this->m_filterBankFilterLast =
+                        std::vector<uint32_t>(this->m_params.m_numFbankBins);
+
+        for (size_t bin = 0; bin < this->m_params.m_numFbankBins; bin++) {
+            float leftMel = melLowFreq + bin * melFreqDelta;
+            float centerMel = melLowFreq + (bin + 1) * melFreqDelta;
+            float rightMel = melLowFreq + (bin + 2) * melFreqDelta;
+
+            uint32_t firstIndex = 0;
+            uint32_t lastIndex = 0;
+            bool firstIndexFound = false;
+            const float normaliser = this->GetMelFilterBankNormaliser(leftMel, rightMel, this->m_params.m_useHtkMethod);
+
+            for (size_t i = 0; i < numFftBins; i++) {
+                float freq = (fftBinWidth * i);  /* Center freq of this fft bin. */
+                float mel = MFCC::MelScale(freq, this->m_params.m_useHtkMethod);
+                thisBin[i] = 0.0;
+
+                if (mel > leftMel && mel < rightMel) {
+                    float weight;
+                    if (mel <= centerMel) {
+                        weight = (mel - leftMel) / (centerMel - leftMel);
+                    } else {
+                        weight = (rightMel - mel) / (rightMel - centerMel);
+                    }
+
+                    thisBin[i] = weight * normaliser;
+                    if (!firstIndexFound) {
+                        firstIndex = i;
+                        firstIndexFound = true;
+                    }
+                    lastIndex = i;
+                }
+            }
+
+            this->m_filterBankFilterFirst[bin] = firstIndex;
+            this->m_filterBankFilterLast[bin] = lastIndex;
+
+            /* Copy the part we care about. */
+            for (uint32_t i = firstIndex; i <= lastIndex; i++) {
+                melFilterBank[bin].push_back(thisBin[i]);
+            }
+        }
+
+        return melFilterBank;
+    }
+
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
diff --git a/source/application/api/common/source/Model.cc b/source/application/api/common/source/Model.cc
new file mode 100644
index 0000000..f1ac91d
--- /dev/null
+++ b/source/application/api/common/source/Model.cc
@@ -0,0 +1,359 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "Model.hpp"
+#include "log_macros.h"
+
+#include <cinttypes>
+
+/* Initialise the model */
+arm::app::Model::~Model()
+{
+   delete this->m_pInterpreter;
+    /**
+     * No clean-up function available for allocator in TensorFlow Lite Micro yet.
+     **/
+}
+
+arm::app::Model::Model() :
+    m_inited (false),
+    m_type(kTfLiteNoType)
+{
+    this->m_pErrorReporter = tflite::GetMicroErrorReporter();
+}
+
+bool arm::app::Model::Init(uint8_t* tensorArenaAddr,
+                           uint32_t tensorArenaSize,
+                           uint8_t* nnModelAddr,
+                           uint32_t nnModelSize,
+                           tflite::MicroAllocator* allocator)
+{
+    /* Following tf lite micro example:
+     * Map the model into a usable data structure. This doesn't involve any
+     * copying or parsing, it's a very lightweight operation. */
+    debug("loading model from @ 0x%p\n", nnModelAddr);
+    debug("model size: %" PRIu32 " bytes.\n", nnModelSize);
+
+    this->m_pModel = ::tflite::GetModel(nnModelAddr);
+
+    if (this->m_pModel->version() != TFLITE_SCHEMA_VERSION) {
+        this->m_pErrorReporter->Report(
+            "[ERROR] model's schema version %d is not equal "
+            "to supported version %d.",
+            this->m_pModel->version(), TFLITE_SCHEMA_VERSION);
+        return false;
+    }
+
+    this->m_modelAddr = nnModelAddr;
+    this->m_modelSize = nnModelSize;
+
+    /* Pull in only the operation implementations we need.
+     * This relies on a complete list of all the ops needed by this graph.
+     * An easier approach is to just use the AllOpsResolver, but this will
+     * incur some penalty in code space for op implementations that are not
+     * needed by this graph.
+     * static ::tflite::ops::micro::AllOpsResolver resolver; */
+    /* NOLINTNEXTLINE(runtime-global-variables) */
+    debug("loading op resolver\n");
+
+    this->EnlistOperations();
+
+    /* Create allocator instance, if it doesn't exist */
+    this->m_pAllocator = allocator;
+    if (!this->m_pAllocator) {
+        /* Create an allocator instance */
+        info("Creating allocator using tensor arena at 0x%p\n", tensorArenaAddr);
+
+        this->m_pAllocator = tflite::MicroAllocator::Create(
+                                        tensorArenaAddr,
+                                        tensorArenaSize,
+                                        this->m_pErrorReporter);
+
+        if (!this->m_pAllocator) {
+            printf_err("Failed to create allocator\n");
+            return false;
+        }
+        debug("Created new allocator @ 0x%p\n", this->m_pAllocator);
+    } else {
+        debug("Using existing allocator @ 0x%p\n", this->m_pAllocator);
+    }
+
+    this->m_pInterpreter = new ::tflite::MicroInterpreter(
+        this->m_pModel, this->GetOpResolver(),
+        this->m_pAllocator, this->m_pErrorReporter);
+
+    if (!this->m_pInterpreter) {
+        printf_err("Failed to allocate interpreter\n");
+        return false;
+    }
+
+    /* Allocate memory from the tensor_arena for the model's tensors. */
+    info("Allocating tensors\n");
+    TfLiteStatus allocate_status = this->m_pInterpreter->AllocateTensors();
+
+    if (allocate_status != kTfLiteOk) {
+        printf_err("tensor allocation failed!\n");
+        delete this->m_pInterpreter;
+        return false;
+    }
+
+    /* Get information about the memory area to use for the model's input. */
+    this->m_input.resize(this->GetNumInputs());
+    for (size_t inIndex = 0; inIndex < this->GetNumInputs(); inIndex++)
+        this->m_input[inIndex] = this->m_pInterpreter->input(inIndex);
+
+    this->m_output.resize(this->GetNumOutputs());
+    for (size_t outIndex = 0; outIndex < this->GetNumOutputs(); outIndex++)
+        this->m_output[outIndex] = this->m_pInterpreter->output(outIndex);
+
+    if (this->m_input.empty() || this->m_output.empty()) {
+        printf_err("failed to get tensors\n");
+        return false;
+    } else {
+        this->m_type = this->m_input[0]->type;  /* Input 0 should be the main input */
+
+        /* Clear the input & output tensors */
+        for (size_t inIndex = 0; inIndex < this->GetNumInputs(); inIndex++) {
+            std::memset(this->m_input[inIndex]->data.data, 0, this->m_input[inIndex]->bytes);
+        }
+        for (size_t outIndex = 0; outIndex < this->GetNumOutputs(); outIndex++) {
+            std::memset(this->m_output[outIndex]->data.data, 0, this->m_output[outIndex]->bytes);
+        }
+
+        this->LogInterpreterInfo();
+    }
+
+    this->m_inited = true;
+    return true;
+}
+
+tflite::MicroAllocator* arm::app::Model::GetAllocator()
+{
+    if (this->IsInited()) {
+        return this->m_pAllocator;
+    }
+    return nullptr;
+}
+
+void arm::app::Model::LogTensorInfo(TfLiteTensor* tensor)
+{
+    if (!tensor) {
+        printf_err("Invalid tensor\n");
+        assert(tensor);
+        return;
+    }
+
+    debug("\ttensor is assigned to 0x%p\n", tensor);
+    info("\ttensor type is %s\n", TfLiteTypeGetName(tensor->type));
+    info("\ttensor occupies %zu bytes with dimensions\n",
+         tensor->bytes);
+    for (int i = 0 ; i < tensor->dims->size; ++i) {
+        info ("\t\t%d: %3d\n", i, tensor->dims->data[i]);
+    }
+
+    TfLiteQuantization quant = tensor->quantization;
+    if (kTfLiteAffineQuantization == quant.type) {
+        auto* quantParams = (TfLiteAffineQuantization*)quant.params;
+        info("Quant dimension: %" PRIi32 "\n", quantParams->quantized_dimension);
+        for (int i = 0; i < quantParams->scale->size; ++i) {
+            info("Scale[%d] = %f\n", i, quantParams->scale->data[i]);
+        }
+        for (int i = 0; i < quantParams->zero_point->size; ++i) {
+            info("ZeroPoint[%d] = %d\n", i, quantParams->zero_point->data[i]);
+        }
+    }
+}
+
+void arm::app::Model::LogInterpreterInfo()
+{
+    if (!this->m_pInterpreter) {
+        printf_err("Invalid interpreter\n");
+        return;
+    }
+
+    info("Model INPUT tensors: \n");
+    for (auto input : this->m_input) {
+        this->LogTensorInfo(input);
+    }
+
+    info("Model OUTPUT tensors: \n");
+    for (auto output : this->m_output) {
+        this->LogTensorInfo(output);
+    }
+
+    info("Activation buffer (a.k.a tensor arena) size used: %zu\n",
+        this->m_pInterpreter->arena_used_bytes());
+
+    /* We expect there to be only one subgraph. */
+    const uint32_t nOperators = tflite::NumSubgraphOperators(this->m_pModel, 0);
+    info("Number of operators: %" PRIu32 "\n", nOperators);
+
+    const tflite::SubGraph* subgraph = this->m_pModel->subgraphs()->Get(0);
+
+    auto* opcodes = this->m_pModel->operator_codes();
+
+    /* For each operator, display registration information. */
+    for (size_t i = 0 ; i < nOperators; ++i) {
+        const tflite::Operator* op = subgraph->operators()->Get(i);
+        const tflite::OperatorCode* opcode = opcodes->Get(op->opcode_index());
+        const TfLiteRegistration* reg = nullptr;
+
+        tflite::GetRegistrationFromOpCode(opcode, this->GetOpResolver(),
+                                          this->m_pErrorReporter, &reg);
+        std::string opName;
+
+        if (reg) {
+            if (tflite::BuiltinOperator_CUSTOM == reg->builtin_code) {
+                opName = std::string(reg->custom_name);
+            } else {
+                opName = std::string(EnumNameBuiltinOperator(
+                            tflite::BuiltinOperator(reg->builtin_code)));
+            }
+        }
+        info("\tOperator %zu: %s\n", i, opName.c_str());
+    }
+}
+
+bool arm::app::Model::IsInited() const
+{
+    return this->m_inited;
+}
+
+bool arm::app::Model::IsDataSigned() const
+{
+    return this->GetType() == kTfLiteInt8;
+}
+
+bool arm::app::Model::ContainsEthosUOperator() const
+{
+    /* We expect there to be only one subgraph. */
+    const uint32_t nOperators = tflite::NumSubgraphOperators(this->m_pModel, 0);
+    const tflite::SubGraph* subgraph = this->m_pModel->subgraphs()->Get(0);
+    const auto* opcodes = this->m_pModel->operator_codes();
+
+    /* check for custom operators */
+    for (size_t i = 0; (i < nOperators); ++i)
+    {
+        const tflite::Operator* op = subgraph->operators()->Get(i);
+        const tflite::OperatorCode* opcode = opcodes->Get(op->opcode_index());
+
+        auto builtin_code = tflite::GetBuiltinCode(opcode);
+        if ((builtin_code == tflite::BuiltinOperator_CUSTOM) &&
+            ( nullptr != opcode->custom_code()) &&
+            ( "ethos-u" == std::string(opcode->custom_code()->c_str())))
+        {
+            return true;
+        }
+    }
+    return false;
+}
+
+bool arm::app::Model::RunInference()
+{
+    bool inference_state = false;
+    if (this->m_pModel && this->m_pInterpreter) {
+        if (kTfLiteOk != this->m_pInterpreter->Invoke()) {
+            printf_err("Invoke failed.\n");
+        } else {
+            inference_state = true;
+        }
+    } else {
+        printf_err("Error: No interpreter!\n");
+    }
+    return inference_state;
+}
+
+TfLiteTensor* arm::app::Model::GetInputTensor(size_t index) const
+{
+    if (index < this->GetNumInputs()) {
+        return this->m_input.at(index);
+    }
+    return nullptr;
+}
+
+TfLiteTensor* arm::app::Model::GetOutputTensor(size_t index) const
+{
+    if (index < this->GetNumOutputs()) {
+        return this->m_output.at(index);
+    }
+    return nullptr;
+}
+
+size_t arm::app::Model::GetNumInputs() const
+{
+    if (this->m_pModel && this->m_pInterpreter) {
+        return this->m_pInterpreter->inputs_size();
+    }
+    return 0;
+}
+
+size_t arm::app::Model::GetNumOutputs() const
+{
+    if (this->m_pModel && this->m_pInterpreter) {
+        return this->m_pInterpreter->outputs_size();
+    }
+    return 0;
+}
+
+
+TfLiteType arm::app::Model::GetType() const
+{
+    return this->m_type;
+}
+
+TfLiteIntArray* arm::app::Model::GetInputShape(size_t index) const
+{
+    if (index < this->GetNumInputs()) {
+        return this->m_input.at(index)->dims;
+    }
+    return nullptr;
+}
+
+TfLiteIntArray* arm::app::Model::GetOutputShape(size_t index) const
+{
+    if (index < this->GetNumOutputs()) {
+        return this->m_output.at(index)->dims;
+    }
+    return nullptr;
+}
+
+bool arm::app::Model::ShowModelInfoHandler()
+{
+    if (!this->IsInited()) {
+        printf_err("Model is not initialised! Terminating processing.\n");
+        return false;
+    }
+
+    PrintTensorFlowVersion();
+    info("Model address: 0x%p", this->ModelPointer());
+    info("Model size:      %" PRIu32 " bytes.", this->ModelSize());
+    info("Model info:\n");
+    this->LogInterpreterInfo();
+
+    info("The model is optimised for Ethos-U NPU: %s.\n", this->ContainsEthosUOperator()? "yes": "no");
+
+    return true;
+}
+
+const uint8_t* arm::app::Model::ModelPointer()
+{
+    return this->m_modelAddr;
+}
+
+uint32_t arm::app::Model::ModelSize()
+{
+    return this->m_modelSize;
+}
diff --git a/source/application/api/common/source/TensorFlowLiteMicro.cc b/source/application/api/common/source/TensorFlowLiteMicro.cc
new file mode 100644
index 0000000..8738e5c
--- /dev/null
+++ b/source/application/api/common/source/TensorFlowLiteMicro.cc
@@ -0,0 +1,46 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "TensorFlowLiteMicro.hpp"
+
+void PrintTensorFlowVersion()
+{}
+
+arm::app::QuantParams arm::app::GetTensorQuantParams(TfLiteTensor* tensor)
+{
+    arm::app::QuantParams params;
+    if (kTfLiteAffineQuantization == tensor->quantization.type) {
+        auto* quantParams = (TfLiteAffineQuantization*) (tensor->quantization.params);
+        if (quantParams && 0 == quantParams->quantized_dimension) {
+            if (quantParams->scale->size) {
+                params.scale = quantParams->scale->data[0];
+            }
+            if (quantParams->zero_point->size) {
+                params.offset = quantParams->zero_point->data[0];
+            }
+        } else if (tensor->params.scale != 0.0) {
+            /* Legacy tensorflow quantisation parameters */
+            params.scale = tensor->params.scale;
+            params.offset = tensor->params.zero_point;
+        }
+    }
+    return params;
+}
+
+extern "C" void DebugLog(const char* s)
+{
+    puts(s);
+}
diff --git a/source/application/api/use_case/ad/CMakeLists.txt b/source/application/api/use_case/ad/CMakeLists.txt
new file mode 100644
index 0000000..224816f
--- /dev/null
+++ b/source/application/api/use_case/ad/CMakeLists.txt
@@ -0,0 +1,41 @@
+#----------------------------------------------------------------------------
+#  Copyright (c) 2022 Arm Limited. All rights reserved.
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#----------------------------------------------------------------------------
+#########################################################
+#             ANOMALY DETECTION API library             #
+#########################################################
+cmake_minimum_required(VERSION 3.15.6)
+
+set(AD_API_TARGET ad_api)
+project(${AD_API_TARGET}
+        DESCRIPTION     "Anomaly detection use case API library"
+        LANGUAGES       C CXX)
+
+# Create static library
+add_library(${AD_API_TARGET} STATIC
+    src/AdModel.cc
+    src/AdProcessing.cc
+    src/AdMelSpectrogram.cc
+    src/MelSpectrogram.cc)
+
+target_include_directories(${AD_API_TARGET} PUBLIC include)
+
+target_link_libraries(${AD_API_TARGET} PUBLIC common_api)
+
+message(STATUS "*******************************************************")
+message(STATUS "Library                                : " ${AD_API_TARGET})
+message(STATUS "CMAKE_SYSTEM_PROCESSOR                 : " ${CMAKE_SYSTEM_PROCESSOR})
+message(STATUS "*******************************************************")
diff --git a/source/application/api/use_case/ad/include/AdMelSpectrogram.hpp b/source/application/api/use_case/ad/include/AdMelSpectrogram.hpp
new file mode 100644
index 0000000..05c5bfc
--- /dev/null
+++ b/source/application/api/use_case/ad/include/AdMelSpectrogram.hpp
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef ADMELSPECTROGRAM_HPP
+#define ADMELSPECTROGRAM_HPP
+
+#include "MelSpectrogram.hpp"
+
+namespace arm {
+namespace app {
+namespace audio {
+
+    /* Class to provide anomaly detection specific Mel Spectrogram calculation requirements */
+    class AdMelSpectrogram : public MelSpectrogram {
+
+    public:
+        static constexpr uint32_t  ms_defaultSamplingFreq = 16000;
+        static constexpr uint32_t  ms_defaultNumFbankBins =    64;
+        static constexpr uint32_t  ms_defaultMelLoFreq    =     0;
+        static constexpr uint32_t  ms_defaultMelHiFreq    =  8000;
+        static constexpr bool      ms_defaultUseHtkMethod = false;
+
+        explicit AdMelSpectrogram(const size_t frameLen)
+                :  MelSpectrogram(MelSpecParams(
+                ms_defaultSamplingFreq, ms_defaultNumFbankBins,
+                ms_defaultMelLoFreq, ms_defaultMelHiFreq,
+                frameLen, ms_defaultUseHtkMethod))
+        {}
+
+        AdMelSpectrogram()  = delete;
+        ~AdMelSpectrogram() = default;
+
+    protected:
+
+        /**
+         * @brief       Overrides base class implementation of this function.
+         * @param[in]   fftVec                  Vector populated with FFT magnitudes
+         * @param[in]   melFilterBank           2D Vector with filter bank weights
+         * @param[in]   filterBankFilterFirst   Vector containing the first indices of filter bank
+         *                                      to be used for each bin.
+         * @param[in]   filterBankFilterLast    Vector containing the last indices of filter bank
+         *                                      to be used for each bin.
+         * @param[out]  melEnergies             Pre-allocated vector of MEL energies to be
+         *                                      populated.
+         * @return      true if successful, false otherwise
+         */
+        virtual bool ApplyMelFilterBank(
+                std::vector<float>&                 fftVec,
+                std::vector<std::vector<float>>&    melFilterBank,
+                std::vector<uint32_t>&               filterBankFilterFirst,
+                std::vector<uint32_t>&               filterBankFilterLast,
+                std::vector<float>&                 melEnergies) override;
+
+        /**
+         * @brief       Override for the base class implementation convert mel
+         *              energies to logarithmic scale. The difference from
+         *              default behaviour is that the power is converted to dB
+         *              and subsequently clamped.
+         * @param[in,out]   melEnergies - 1D vector of Mel energies
+         **/
+        virtual void ConvertToLogarithmicScale(std::vector<float>& melEnergies) override;
+
+        /**
+         * @brief       Given the low and high Mel values, get the normaliser
+         *              for weights to be applied when populating the filter
+         *              bank. Override for the base class implementation.
+         * @param[in]   leftMel - low Mel frequency value
+         * @param[in]   rightMel - high Mel frequency value
+         * @param[in]   useHTKMethod - bool to signal if HTK method is to be
+         *              used for calculation
+         * @return      Return float value to be applied 
+         *              when populating the filter bank.
+         */
+        virtual float GetMelFilterBankNormaliser(
+                const float&   leftMel,
+                const float&   rightMel,
+                const bool     useHTKMethod) override;
+    };
+
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* ADMELSPECTROGRAM_HPP */
diff --git a/source/application/api/use_case/ad/include/AdModel.hpp b/source/application/api/use_case/ad/include/AdModel.hpp
new file mode 100644
index 0000000..0436a89
--- /dev/null
+++ b/source/application/api/use_case/ad/include/AdModel.hpp
@@ -0,0 +1,55 @@
+/*
+ * Copyright (c) 2021-2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef AD_MODEL_HPP
+#define AD_MODEL_HPP
+
+#include "Model.hpp"
+
+extern const int g_FrameLength;
+extern const int g_FrameStride;
+extern const float g_ScoreThreshold;
+extern const float g_TrainingMean;
+
+namespace arm {
+namespace app {
+
+    class AdModel : public Model {
+
+    public:
+        /* Indices for the expected model - based on input tensor shape */
+        static constexpr uint32_t ms_inputRowsIdx = 1;
+        static constexpr uint32_t ms_inputColsIdx = 2;
+
+    protected:
+        /** @brief   Gets the reference to op resolver interface class */
+        const tflite::MicroOpResolver& GetOpResolver() override;
+
+        /** @brief   Adds operations to the op resolver instance */
+        bool EnlistOperations() override;
+
+    private:
+        /* Maximum number of individual operations that can be enlisted */
+        static constexpr int ms_maxOpCnt = 6;
+
+        /* A mutable op resolver instance */
+        tflite::MicroMutableOpResolver<ms_maxOpCnt> m_opResolver;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* AD_MODEL_HPP */
diff --git a/source/application/api/use_case/ad/include/AdProcessing.hpp b/source/application/api/use_case/ad/include/AdProcessing.hpp
new file mode 100644
index 0000000..abee75e
--- /dev/null
+++ b/source/application/api/use_case/ad/include/AdProcessing.hpp
@@ -0,0 +1,231 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef AD_PROCESSING_HPP
+#define AD_PROCESSING_HPP
+
+#include "BaseProcessing.hpp"
+#include "TensorFlowLiteMicro.hpp"
+#include "AudioUtils.hpp"
+#include "AdMelSpectrogram.hpp"
+#include "log_macros.h"
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   Pre-processing class for anomaly detection use case.
+     *          Implements methods declared by BasePreProcess and anything else needed
+     *          to populate input tensors ready for inference.
+     */
+    class AdPreProcess : public BasePreProcess {
+
+    public:
+        /**
+         * @brief Constructor for AdPreProcess class objects
+         * @param[in] inputTensor  input tensor pointer from the tensor arena.
+         * @param[in] melSpectrogramFrameLen MEL spectrogram's frame length
+         * @param[in] melSpectrogramFrameStride MEL spectrogram's frame stride
+         * @param[in] adModelTrainingMean Training mean for the Anomaly detection model being used.
+         */
+        explicit AdPreProcess(TfLiteTensor* inputTensor,
+                              uint32_t melSpectrogramFrameLen,
+                              uint32_t melSpectrogramFrameStride,
+                              float adModelTrainingMean);
+
+        ~AdPreProcess() = default;
+
+        /**
+         * @brief Function to invoke pre-processing and populate the input vector
+         * @param input pointer to input data. For anomaly detection, this is the pointer to
+         *              the audio data.
+         * @param inputSize Size of the data being passed in for pre-processing.
+         * @return True if successful, false otherwise.
+         */
+        bool DoPreProcess(const void* input, size_t inputSize) override;
+
+        /**
+         * @brief Getter function for audio window size computed when constructing
+         *        the class object.
+         * @return Audio window size as 32 bit unsigned integer.
+         */
+        uint32_t GetAudioWindowSize();
+
+        /**
+         * @brief Getter function for audio window stride computed when constructing
+         *        the class object.
+         * @return Audio window stride as 32 bit unsigned integer.
+         */
+        uint32_t GetAudioDataStride();
+
+        /**
+         * @brief Setter function for current audio index. This is only used for evaluating
+         *        if previously computed features can be re-used from cache.
+         */
+        void SetAudioWindowIndex(uint32_t idx);
+
+    private:
+        bool        m_validInstance{false}; /**< Indicates the current object is valid. */
+        uint32_t    m_melSpectrogramFrameLen{}; /**< MEL spectrogram's window frame length */
+        uint32_t    m_melSpectrogramFrameStride{}; /**< MEL spectrogram's window frame stride */
+        uint8_t     m_inputResizeScale{}; /**< Downscaling factor for the MEL energy matrix. */
+        uint32_t    m_numMelSpecVectorsInAudioStride{};  /**< Number of frames to move across the audio. */
+        uint32_t    m_audioDataWindowSize{}; /**< Audio window size computed based on other parameters. */
+        uint32_t    m_audioDataStride{}; /**< Audio window stride computed. */
+        uint32_t    m_numReusedFeatureVectors{}; /**< Number of MEL vectors that can be re-used */
+        uint32_t    m_audioWindowIndex{}; /**< Current audio window index (from audio's sliding window) */
+
+        audio::SlidingWindow<const int16_t> m_melWindowSlider; /**< Internal MEL spectrogram window slider */
+        audio::AdMelSpectrogram m_melSpec; /**< MEL spectrogram computation object */
+        std::function<void
+            (std::vector<int16_t>&, int, bool, size_t, size_t)> m_featureCalc; /**< Feature calculator object */
+    };
+
+    class AdPostProcess : public BasePostProcess {
+    public:
+        /**
+         * @brief Constructor for AdPostProcess object.
+         * @param[in] outputTensor Output tensor pointer.
+         */
+        explicit AdPostProcess(TfLiteTensor* outputTensor);
+
+        ~AdPostProcess() = default;
+
+        /**
+         * @brief Function to do the post-processing on the output tensor.
+         * @return True if successful, false otherwise.
+         */
+        bool DoPostProcess() override;
+
+        /**
+         * @brief Getter function for an element from the de-quantised output vector.
+         * @param index Index of the element to be retrieved.
+         * @return index represented as a 32 bit floating point number.
+         */
+        float GetOutputValue(uint32_t index);
+
+    private:
+        TfLiteTensor* m_outputTensor{}; /**< Output tensor pointer */
+        std::vector<float> m_dequantizedOutputVec{}; /**< Internal output vector */
+
+        /**
+         * @brief De-quantizes and flattens the output tensor into a vector.
+         * @tparam T template parameter to indicate data type.
+         * @return True if successful, false otherwise.
+         */
+        template<typename T>
+        bool Dequantize()
+        {
+            TfLiteTensor* tensor = this->m_outputTensor;
+            if (tensor == nullptr) {
+                printf_err("Invalid output tensor.\n");
+                return false;
+            }
+            T* tensorData = tflite::GetTensorData<T>(tensor);
+
+            uint32_t totalOutputSize = 1;
+            for (int inputDim = 0; inputDim < tensor->dims->size; inputDim++){
+                totalOutputSize *= tensor->dims->data[inputDim];
+            }
+
+            /* For getting the floating point values, we need quantization parameters */
+            QuantParams quantParams = GetTensorQuantParams(tensor);
+
+            this->m_dequantizedOutputVec = std::vector<float>(totalOutputSize, 0);
+
+            for (size_t i = 0; i < totalOutputSize; ++i) {
+                this->m_dequantizedOutputVec[i] = quantParams.scale * (tensorData[i] - quantParams.offset);
+            }
+
+            return true;
+        }
+    };
+
+    /* Templated instances available: */
+    template bool AdPostProcess::Dequantize<int8_t>();
+
+    /**
+     * @brief Generic feature calculator factory.
+     *
+     * Returns lambda function to compute features using features cache.
+     * Real features math is done by a lambda function provided as a parameter.
+     * Features are written to input tensor memory.
+     *
+     * @tparam T            feature vector type.
+     * @param inputTensor   model input tensor pointer.
+     * @param cacheSize     number of feature vectors to cache. Defined by the sliding window overlap.
+     * @param compute       features calculator function.
+     * @return              lambda function to compute features.
+     */
+    template<class T>
+    std::function<void (std::vector<int16_t>&, size_t, bool, size_t, size_t)>
+    FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize,
+                std::function<std::vector<T> (std::vector<int16_t>& )> compute)
+    {
+        /* Feature cache to be captured by lambda function*/
+        static std::vector<std::vector<T>> featureCache = std::vector<std::vector<T>>(cacheSize);
+
+        return [=](std::vector<int16_t>& audioDataWindow,
+                   size_t index,
+                   bool useCache,
+                   size_t featuresOverlapIndex,
+                   size_t resizeScale)
+        {
+            T* tensorData = tflite::GetTensorData<T>(inputTensor);
+            std::vector<T> features;
+
+            /* Reuse features from cache if cache is ready and sliding windows overlap.
+             * Overlap is in the beginning of sliding window with a size of a feature cache. */
+            if (useCache && index < featureCache.size()) {
+                features = std::move(featureCache[index]);
+            } else {
+                features = std::move(compute(audioDataWindow));
+            }
+            auto size = features.size() / resizeScale;
+            auto sizeBytes = sizeof(T);
+
+            /* Input should be transposed and "resized" by skipping elements. */
+            for (size_t outIndex = 0; outIndex < size; outIndex++) {
+                std::memcpy(tensorData + (outIndex*size) + index, &features[outIndex*resizeScale], sizeBytes);
+            }
+
+            /* Start renewing cache as soon iteration goes out of the windows overlap. */
+            if (index >= featuresOverlapIndex / resizeScale) {
+                featureCache[index - featuresOverlapIndex / resizeScale] = std::move(features);
+            }
+        };
+    }
+
+    template std::function<void (std::vector<int16_t>&, size_t , bool, size_t, size_t)>
+    FeatureCalc<int8_t>(TfLiteTensor* inputTensor,
+                        size_t cacheSize,
+                        std::function<std::vector<int8_t> (std::vector<int16_t>&)> compute);
+
+    template std::function<void(std::vector<int16_t>&, size_t, bool, size_t, size_t)>
+    FeatureCalc<float>(TfLiteTensor *inputTensor,
+                       size_t cacheSize,
+                       std::function<std::vector<float>(std::vector<int16_t>&)> compute);
+
+    std::function<void (std::vector<int16_t>&, int, bool, size_t, size_t)>
+    GetFeatureCalculator(audio::AdMelSpectrogram& melSpec,
+                         TfLiteTensor* inputTensor,
+                         size_t cacheSize,
+                         float trainingMean);
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* AD_PROCESSING_HPP */
diff --git a/source/application/api/use_case/ad/include/MelSpectrogram.hpp b/source/application/api/use_case/ad/include/MelSpectrogram.hpp
new file mode 100644
index 0000000..d3ea3f7
--- /dev/null
+++ b/source/application/api/use_case/ad/include/MelSpectrogram.hpp
@@ -0,0 +1,234 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MELSPECTROGRAM_HPP
+#define MELSPECTROGRAM_HPP
+
+#include "PlatformMath.hpp"
+
+#include <vector>
+#include <cstdint>
+#include <cmath>
+#include <limits>
+#include <string>
+
+namespace arm {
+namespace app {
+namespace audio {
+
+    /* Mel Spectrogram consolidated parameters */
+    class MelSpecParams {
+    public:
+        float       m_samplingFreq;
+        uint32_t    m_numFbankBins;
+        float       m_melLoFreq;
+        float       m_melHiFreq;
+        uint32_t    m_frameLen;
+        uint32_t    m_frameLenPadded;
+        bool        m_useHtkMethod;
+
+        /** @brief  Constructor */
+        MelSpecParams(const float samplingFreq, const uint32_t numFbankBins,
+                      const float melLoFreq, const float melHiFreq,
+                      const uint32_t frameLen, const bool useHtkMethod);
+
+        MelSpecParams()  = delete;
+        ~MelSpecParams() = default;
+
+        /** @brief  String representation of parameters */
+        std::string Str() const;
+    };
+
+    /**
+     * @brief   Class for Mel Spectrogram feature extraction.
+     *          Based on https://github.com/ARM-software/ML-KWS-for-MCU/blob/master/Deployment/Source/MFCC/mfcc.cpp
+     *          This class is designed to be generic and self-sufficient but
+     *          certain calculation routines can be overridden to accommodate
+     *          use-case specific requirements.
+     */
+    class MelSpectrogram {
+
+    public:
+        /**
+        * @brief        Extract Mel Spectrogram for one single small frame of
+        *               audio data e.g. 640 samples.
+        * @param[in]    audioData       Vector of audio samples to calculate
+        *               features for.
+        * @param[in]    trainingMean    Value to subtract from the the computed mel spectrogram, default 0.
+        * @return       Vector of extracted Mel Spectrogram features.
+        **/
+        std::vector<float> ComputeMelSpec(const std::vector<int16_t>& audioData, float trainingMean = 0);
+
+        /**
+         * @brief       Constructor
+         * @param[in]   params   Mel Spectrogram parameters
+        */
+        explicit MelSpectrogram(const MelSpecParams& params);
+
+        MelSpectrogram() = delete;
+        ~MelSpectrogram() = default;
+
+        /** @brief  Initialise */
+        void Init();
+
+        /**
+         * @brief        Extract Mel Spectrogram features and quantise for one single small
+         *               frame of audio data e.g. 640 samples.
+         * @param[in]    audioData      Vector of audio samples to calculate
+         *               features for.
+         * @param[in]    quantScale     quantisation scale.
+         * @param[in]    quantOffset    quantisation offset.
+         * @param[in]    trainingMean   training mean.
+         * @return       Vector of extracted quantised Mel Spectrogram features.
+         **/
+        template<typename T>
+        std::vector<T> MelSpecComputeQuant(const std::vector<int16_t>& audioData,
+                                           const float quantScale,
+                                           const int quantOffset,
+                                           float trainingMean = 0)
+        {
+            this->ComputeMelSpec(audioData, trainingMean);
+            float minVal = std::numeric_limits<T>::min();
+            float maxVal = std::numeric_limits<T>::max();
+
+            std::vector<T> melSpecOut(this->m_params.m_numFbankBins);
+            const size_t numFbankBins = this->m_params.m_numFbankBins;
+
+            /* Quantize to T. */
+            for (size_t k = 0; k < numFbankBins; ++k) {
+                auto quantizedEnergy = std::round(((this->m_melEnergies[k]) / quantScale) + quantOffset);
+                melSpecOut[k] = static_cast<T>(std::min<float>(std::max<float>(quantizedEnergy, minVal), maxVal));
+            }
+
+            return melSpecOut;
+        }
+
+        /* Constants */
+        static constexpr float ms_logStep = /*logf(6.4)*/ 1.8562979903656 / 27.0;
+        static constexpr float ms_freqStep = 200.0 / 3;
+        static constexpr float ms_minLogHz = 1000.0;
+        static constexpr float ms_minLogMel = ms_minLogHz / ms_freqStep;
+
+    protected:
+        /**
+         * @brief       Project input frequency to Mel Scale.
+         * @param[in]   freq          input frequency in floating point
+         * @param[in]   useHTKMethod  bool to signal if HTK method is to be
+         *                            used for calculation
+         * @return      Mel transformed frequency in floating point
+         **/
+        static float MelScale(const float    freq,
+                              const bool     useHTKMethod = true);
+
+        /**
+         * @brief       Inverse Mel transform - convert MEL warped frequency
+         *              back to normal frequency
+         * @param[in]   melFreq          Mel frequency in floating point
+         * @param[in]   useHTKMethod  bool to signal if HTK method is to be
+         *                            used for calculation
+         * @return      Real world frequency in floating point
+         **/
+        static float InverseMelScale(const float melFreq,
+                                     const bool  useHTKMethod = true);
+
+        /**
+         * @brief       Populates MEL energies after applying the MEL filter
+         *              bank weights and adding them up to be placed into
+         *              bins, according to the filter bank's first and last
+         *              indices (pre-computed for each filter bank element
+         *              by CreateMelFilterBank function).
+         * @param[in]   fftVec                  Vector populated with FFT magnitudes
+         * @param[in]   melFilterBank           2D Vector with filter bank weights
+         * @param[in]   filterBankFilterFirst   Vector containing the first indices of filter bank
+         *                                      to be used for each bin.
+         * @param[in]   filterBankFilterLast    Vector containing the last indices of filter bank
+         *                                      to be used for each bin.
+         * @param[out]  melEnergies             Pre-allocated vector of MEL energies to be
+         *                                      populated.
+         * @return      true if successful, false otherwise
+         */
+        virtual bool ApplyMelFilterBank(
+                std::vector<float>&                 fftVec,
+                std::vector<std::vector<float>>&    melFilterBank,
+                std::vector<uint32_t>&               filterBankFilterFirst,
+                std::vector<uint32_t>&               filterBankFilterLast,
+                std::vector<float>&                 melEnergies);
+
+        /**
+         * @brief           Converts the Mel energies for logarithmic scale
+         * @param[in,out]   melEnergies 1D vector of Mel energies
+         **/
+        virtual void ConvertToLogarithmicScale(std::vector<float>& melEnergies);
+
+        /**
+         * @brief       Given the low and high Mel values, get the normaliser
+         *              for weights to be applied when populating the filter
+         *              bank.
+         * @param[in]   leftMel      low Mel frequency value
+         * @param[in]   rightMel     high Mel frequency value
+         * @param[in]   useHTKMethod bool to signal if HTK method is to be
+         *                           used for calculation
+         * @return      Return float value to be applied 
+         *              when populating the filter bank.
+         */
+        virtual float GetMelFilterBankNormaliser(
+                const float&   leftMel,
+                const float&   rightMel,
+                const bool     useHTKMethod);
+
+    private:
+        MelSpecParams                   m_params;
+        std::vector<float>              m_frame;
+        std::vector<float>              m_buffer;
+        std::vector<float>              m_melEnergies;
+        std::vector<float>              m_windowFunc;
+        std::vector<std::vector<float>> m_melFilterBank;
+        std::vector<uint32_t>            m_filterBankFilterFirst;
+        std::vector<uint32_t>            m_filterBankFilterLast;
+        bool                            m_filterBankInitialised;
+        arm::app::math::FftInstance     m_fftInstance;
+
+        /**
+         * @brief       Initialises the filter banks.
+         **/
+        void InitMelFilterBank();
+
+        /**
+         * @brief       Signals whether the instance of MelSpectrogram has had its
+         *              required buffers initialised
+         * @return      True if initialised, false otherwise
+         **/
+        bool IsMelFilterBankInited() const;
+
+        /**
+         * @brief       Create mel filter banks for Mel Spectrogram calculation.
+         * @return      2D vector of floats
+         **/
+        std::vector<std::vector<float>> CreateMelFilterBank();
+
+        /**
+         * @brief       Computes the magnitude from an interleaved complex array
+         **/
+        void ConvertToPowerSpectrum();
+
+    };
+
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
+
+
+#endif /* MELSPECTROGRAM_HPP */
diff --git a/source/application/api/use_case/ad/src/AdMelSpectrogram.cc b/source/application/api/use_case/ad/src/AdMelSpectrogram.cc
new file mode 100644
index 0000000..14b9323
--- /dev/null
+++ b/source/application/api/use_case/ad/src/AdMelSpectrogram.cc
@@ -0,0 +1,93 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "AdMelSpectrogram.hpp"
+#include "PlatformMath.hpp"
+#include "log_macros.h"
+
+#include <cfloat>
+
+namespace arm {
+namespace app {
+namespace audio {
+
+    bool AdMelSpectrogram::ApplyMelFilterBank(
+            std::vector<float>&                 fftVec,
+            std::vector<std::vector<float>>&    melFilterBank,
+            std::vector<uint32_t>&               filterBankFilterFirst,
+            std::vector<uint32_t>&               filterBankFilterLast,
+            std::vector<float>&                 melEnergies)
+    {
+        const size_t numBanks = melEnergies.size();
+
+        if (numBanks != filterBankFilterFirst.size() ||
+            numBanks != filterBankFilterLast.size()) {
+            printf_err("unexpected filter bank lengths\n");
+            return false;
+        }
+
+        for (size_t bin = 0; bin < numBanks; ++bin) {
+            auto filterBankIter = melFilterBank[bin].begin();
+            auto end = melFilterBank[bin].end();
+            float melEnergy = FLT_MIN; /* Avoid log of zero at later stages. */
+            const uint32_t firstIndex = filterBankFilterFirst[bin];
+            const uint32_t lastIndex = std::min<int32_t>(filterBankFilterLast[bin], fftVec.size() - 1);
+
+            for (uint32_t i = firstIndex; i <= lastIndex && filterBankIter != end; ++i) {
+                melEnergy += (*filterBankIter++ * fftVec[i]);
+            }
+
+            melEnergies[bin] = melEnergy;
+        }
+
+        return true;
+    }
+
+    void AdMelSpectrogram::ConvertToLogarithmicScale(
+            std::vector<float>& melEnergies)
+    {
+        /* Container for natural logarithms of mel energies */
+        std::vector <float> vecLogEnergies(melEnergies.size(), 0.f);
+
+        /* Because we are taking natural logs, we need to multiply by log10(e).
+         * Also, for wav2letter model, we scale our log10 values by 10 */
+        constexpr float multiplier = 10.0 * /* default scalar */
+                                     0.4342944819032518; /* log10f(std::exp(1.0))*/
+
+        /* Take log of the whole vector */
+        math::MathUtils::VecLogarithmF32(melEnergies, vecLogEnergies);
+
+        /* Scale the log values. */
+        for (auto iterM = melEnergies.begin(), iterL = vecLogEnergies.begin();
+             iterM != melEnergies.end() && iterL != vecLogEnergies.end(); ++iterM, ++iterL) {
+
+            *iterM = *iterL * multiplier;
+        }
+    }
+
+    float AdMelSpectrogram::GetMelFilterBankNormaliser(
+            const float&    leftMel,
+            const float&    rightMel,
+            const bool      useHTKMethod)
+    {
+        /* Slaney normalization for mel weights. */
+        return (2.0f / (AdMelSpectrogram::InverseMelScale(rightMel, useHTKMethod) -
+                        AdMelSpectrogram::InverseMelScale(leftMel, useHTKMethod)));
+    }
+
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
diff --git a/source/application/api/use_case/ad/src/AdModel.cc b/source/application/api/use_case/ad/src/AdModel.cc
new file mode 100644
index 0000000..961c260
--- /dev/null
+++ b/source/application/api/use_case/ad/src/AdModel.cc
@@ -0,0 +1,41 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "AdModel.hpp"
+#include "log_macros.h"
+
+const tflite::MicroOpResolver& arm::app::AdModel::GetOpResolver()
+{
+    return this->m_opResolver;
+}
+
+bool arm::app::AdModel::EnlistOperations()
+{
+    this->m_opResolver.AddAveragePool2D();
+    this->m_opResolver.AddConv2D();
+    this->m_opResolver.AddDepthwiseConv2D();
+    this->m_opResolver.AddRelu6();
+    this->m_opResolver.AddReshape();
+
+    if (kTfLiteOk == this->m_opResolver.AddEthosU()) {
+        info("Added %s support to op resolver\n",
+            tflite::GetString_ETHOSU());
+    } else {
+        printf_err("Failed to add Arm NPU support to op resolver.");
+        return false;
+    }
+    return true;
+}
diff --git a/source/application/api/use_case/ad/src/AdProcessing.cc b/source/application/api/use_case/ad/src/AdProcessing.cc
new file mode 100644
index 0000000..fb26a83
--- /dev/null
+++ b/source/application/api/use_case/ad/src/AdProcessing.cc
@@ -0,0 +1,210 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "AdProcessing.hpp"
+
+#include "AdModel.hpp"
+
+namespace arm {
+namespace app {
+
+AdPreProcess::AdPreProcess(TfLiteTensor* inputTensor,
+                           uint32_t melSpectrogramFrameLen,
+                           uint32_t melSpectrogramFrameStride,
+                           float adModelTrainingMean):
+       m_validInstance{false},
+       m_melSpectrogramFrameLen{melSpectrogramFrameLen},
+       m_melSpectrogramFrameStride{melSpectrogramFrameStride},
+        /**< Model is trained on features downsampled 2x */
+       m_inputResizeScale{2},
+        /**< We are choosing to move by 20 frames across the audio for each inference. */
+       m_numMelSpecVectorsInAudioStride{20},
+       m_audioDataStride{m_numMelSpecVectorsInAudioStride * melSpectrogramFrameStride},
+       m_melSpec{melSpectrogramFrameLen}
+{
+    UNUSED(this->m_melSpectrogramFrameStride);
+
+    if (!inputTensor) {
+        printf_err("Invalid input tensor provided to pre-process\n");
+        return;
+    }
+
+    TfLiteIntArray* inputShape = inputTensor->dims;
+
+    if (!inputShape) {
+        printf_err("Invalid input tensor dims\n");
+        return;
+    }
+
+    const uint32_t kNumRows = inputShape->data[AdModel::ms_inputRowsIdx];
+    const uint32_t kNumCols = inputShape->data[AdModel::ms_inputColsIdx];
+
+    /* Deduce the data length required for 1 inference from the network parameters. */
+    this->m_audioDataWindowSize = (((this->m_inputResizeScale * kNumCols) - 1) *
+                                    melSpectrogramFrameStride) +
+                                    melSpectrogramFrameLen;
+    this->m_numReusedFeatureVectors = kNumRows -
+                                      (this->m_numMelSpecVectorsInAudioStride /
+                                       this->m_inputResizeScale);
+    this->m_melSpec.Init();
+
+    /* Creating a Mel Spectrogram sliding window for the data required for 1 inference.
+     * "resizing" done here by multiplying stride by resize scale. */
+    this->m_melWindowSlider = audio::SlidingWindow<const int16_t>(
+            nullptr, /* to be populated later. */
+            this->m_audioDataWindowSize,
+            melSpectrogramFrameLen,
+            melSpectrogramFrameStride * this->m_inputResizeScale);
+
+    /* Construct feature calculation function. */
+    this->m_featureCalc = GetFeatureCalculator(this->m_melSpec, inputTensor,
+                                               this->m_numReusedFeatureVectors,
+                                               adModelTrainingMean);
+    this->m_validInstance = true;
+}
+
+bool AdPreProcess::DoPreProcess(const void* input, size_t inputSize)
+{
+    /* Check that we have a valid instance. */
+    if (!this->m_validInstance) {
+        printf_err("Invalid pre-processor instance\n");
+        return false;
+    }
+
+    /* We expect that we can traverse the size with which the MEL spectrogram
+     * sliding window was initialised with. */
+    if (!input || inputSize < this->m_audioDataWindowSize) {
+        printf_err("Invalid input provided for pre-processing\n");
+        return false;
+    }
+
+    /* We moved to the next window - set the features sliding to the new address. */
+    this->m_melWindowSlider.Reset(static_cast<const int16_t*>(input));
+
+    /* The first window does not have cache ready. */
+    const bool useCache = this->m_audioWindowIndex > 0 && this->m_numReusedFeatureVectors > 0;
+
+    /* Start calculating features inside one audio sliding window. */
+    while (this->m_melWindowSlider.HasNext()) {
+        const int16_t* melSpecWindow = this->m_melWindowSlider.Next();
+        std::vector<int16_t> melSpecAudioData = std::vector<int16_t>(
+                melSpecWindow,
+                melSpecWindow + this->m_melSpectrogramFrameLen);
+
+        /* Compute features for this window and write them to input tensor. */
+        this->m_featureCalc(melSpecAudioData,
+                            this->m_melWindowSlider.Index(),
+                            useCache,
+                            this->m_numMelSpecVectorsInAudioStride,
+                            this->m_inputResizeScale);
+    }
+
+    return true;
+}
+
+uint32_t AdPreProcess::GetAudioWindowSize()
+{
+    return this->m_audioDataWindowSize;
+}
+
+uint32_t AdPreProcess::GetAudioDataStride()
+{
+    return this->m_audioDataStride;
+}
+
+void AdPreProcess::SetAudioWindowIndex(uint32_t idx)
+{
+    this->m_audioWindowIndex = idx;
+}
+
+AdPostProcess::AdPostProcess(TfLiteTensor* outputTensor) :
+    m_outputTensor {outputTensor}
+{}
+
+bool AdPostProcess::DoPostProcess()
+{
+    switch (this->m_outputTensor->type) {
+        case kTfLiteInt8:
+            this->Dequantize<int8_t>();
+            break;
+        default:
+            printf_err("Unsupported tensor type");
+            return false;
+    }
+
+    math::MathUtils::SoftmaxF32(this->m_dequantizedOutputVec);
+    return true;
+}
+
+float AdPostProcess::GetOutputValue(uint32_t index)
+{
+    if (index < this->m_dequantizedOutputVec.size()) {
+        return this->m_dequantizedOutputVec[index];
+    }
+    printf_err("Invalid index for output\n");
+    return 0.0;
+}
+
+std::function<void (std::vector<int16_t>&, int, bool, size_t, size_t)>
+GetFeatureCalculator(audio::AdMelSpectrogram& melSpec,
+                     TfLiteTensor* inputTensor,
+                     size_t cacheSize,
+                     float trainingMean)
+{
+    std::function<void (std::vector<int16_t>&, size_t, bool, size_t, size_t)> melSpecFeatureCalc;
+
+    TfLiteQuantization quant = inputTensor->quantization;
+
+    if (kTfLiteAffineQuantization == quant.type) {
+
+        auto* quantParams = static_cast<TfLiteAffineQuantization*>(quant.params);
+        const float quantScale = quantParams->scale->data[0];
+        const int quantOffset = quantParams->zero_point->data[0];
+
+        switch (inputTensor->type) {
+            case kTfLiteInt8: {
+                melSpecFeatureCalc = FeatureCalc<int8_t>(
+                        inputTensor,
+                        cacheSize,
+                        [=, &melSpec](std::vector<int16_t>& audioDataWindow) {
+                            return melSpec.MelSpecComputeQuant<int8_t>(
+                                    audioDataWindow,
+                                    quantScale,
+                                    quantOffset,
+                                    trainingMean);
+                        }
+                );
+                break;
+            }
+            default:
+            printf_err("Tensor type %s not supported\n", TfLiteTypeGetName(inputTensor->type));
+        }
+    } else {
+        melSpecFeatureCalc = FeatureCalc<float>(
+                inputTensor,
+                cacheSize,
+                [=, &melSpec](
+                        std::vector<int16_t>& audioDataWindow) {
+                    return melSpec.ComputeMelSpec(
+                            audioDataWindow,
+                            trainingMean);
+                });
+    }
+    return melSpecFeatureCalc;
+}
+
+} /* namespace app */
+} /* namespace arm */
diff --git a/source/application/api/use_case/ad/src/MelSpectrogram.cc b/source/application/api/use_case/ad/src/MelSpectrogram.cc
new file mode 100644
index 0000000..ff0c536
--- /dev/null
+++ b/source/application/api/use_case/ad/src/MelSpectrogram.cc
@@ -0,0 +1,316 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "MelSpectrogram.hpp"
+
+#include "PlatformMath.hpp"
+#include "log_macros.h"
+
+#include <cfloat>
+#include <cinttypes>
+
+namespace arm {
+namespace app {
+namespace audio {
+
+    MelSpecParams::MelSpecParams(
+            const float samplingFreq,
+            const uint32_t numFbankBins,
+            const float melLoFreq,
+            const float melHiFreq,
+            const uint32_t frameLen,
+            const bool useHtkMethod):
+            m_samplingFreq(samplingFreq),
+            m_numFbankBins(numFbankBins),
+            m_melLoFreq(melLoFreq),
+            m_melHiFreq(melHiFreq),
+            m_frameLen(frameLen),
+
+            /* Smallest power of 2 >= frame length. */
+            m_frameLenPadded(pow(2, ceil((log(frameLen)/log(2))))),
+            m_useHtkMethod(useHtkMethod)
+    {}
+
+    std::string MelSpecParams::Str() const
+    {
+        char strC[1024];
+        snprintf(strC, sizeof(strC) - 1, "\n   \
+            \n\t Sampling frequency:         %f\
+            \n\t Number of filter banks:     %" PRIu32 "\
+            \n\t Mel frequency limit (low):  %f\
+            \n\t Mel frequency limit (high): %f\
+            \n\t Frame length:               %" PRIu32 "\
+            \n\t Padded frame length:        %" PRIu32 "\
+            \n\t Using HTK for Mel scale:    %s\n",
+            this->m_samplingFreq, this->m_numFbankBins, this->m_melLoFreq,
+            this->m_melHiFreq, this->m_frameLen,
+            this->m_frameLenPadded, this->m_useHtkMethod ? "yes" : "no");
+        return std::string{strC};
+    }
+
+    MelSpectrogram::MelSpectrogram(const MelSpecParams& params):
+            m_params(params),
+            m_filterBankInitialised(false)
+    {
+        this->m_buffer = std::vector<float>(
+                this->m_params.m_frameLenPadded, 0.0);
+        this->m_frame = std::vector<float>(
+                this->m_params.m_frameLenPadded, 0.0);
+        this->m_melEnergies = std::vector<float>(
+                this->m_params.m_numFbankBins, 0.0);
+
+        this->m_windowFunc = std::vector<float>(this->m_params.m_frameLen);
+        const auto multiplier = static_cast<float>(2 * M_PI / this->m_params.m_frameLen);
+
+        /* Create window function. */
+        for (size_t i = 0; i < this->m_params.m_frameLen; ++i) {
+            this->m_windowFunc[i] = (0.5 - (0.5 *
+                                             math::MathUtils::CosineF32(static_cast<float>(i) * multiplier)));
+        }
+
+        math::MathUtils::FftInitF32(this->m_params.m_frameLenPadded, this->m_fftInstance);
+        debug("Instantiated Mel Spectrogram object: %s\n", this->m_params.Str().c_str());
+    }
+
+    void MelSpectrogram::Init()
+    {
+        this->InitMelFilterBank();
+    }
+
+    float MelSpectrogram::MelScale(const float freq, const bool useHTKMethod)
+    {
+        if (useHTKMethod) {
+            return 1127.0f * logf (1.0f + freq / 700.0f);
+        } else {
+            /* Slaney formula for mel scale. */
+            float mel = freq / ms_freqStep;
+
+            if (freq >= ms_minLogHz) {
+                mel = ms_minLogMel + logf(freq / ms_minLogHz) / ms_logStep;
+            }
+            return mel;
+        }
+    }
+
+    float MelSpectrogram::InverseMelScale(const float melFreq, const bool useHTKMethod)
+    {
+        if (useHTKMethod) {
+            return 700.0f * (expf (melFreq / 1127.0f) - 1.0f);
+        } else {
+            /* Slaney formula for inverse mel scale. */
+            float freq = ms_freqStep * melFreq;
+
+            if (melFreq >= ms_minLogMel) {
+                freq = ms_minLogHz * expf(ms_logStep * (melFreq - ms_minLogMel));
+            }
+            return freq;
+        }
+    }
+
+    bool MelSpectrogram::ApplyMelFilterBank(
+            std::vector<float>&                 fftVec,
+            std::vector<std::vector<float>>&    melFilterBank,
+            std::vector<uint32_t>&               filterBankFilterFirst,
+            std::vector<uint32_t>&               filterBankFilterLast,
+            std::vector<float>&                 melEnergies)
+    {
+        const size_t numBanks = melEnergies.size();
+
+        if (numBanks != filterBankFilterFirst.size() ||
+            numBanks != filterBankFilterLast.size()) {
+            printf_err("unexpected filter bank lengths\n");
+            return false;
+        }
+
+        for (size_t bin = 0; bin < numBanks; ++bin) {
+            auto filterBankIter = melFilterBank[bin].begin();
+            auto end = melFilterBank[bin].end();
+            float melEnergy = FLT_MIN; /* Avoid log of zero at later stages */
+            const uint32_t firstIndex = filterBankFilterFirst[bin];
+            const uint32_t lastIndex = std::min<int32_t>(filterBankFilterLast[bin], fftVec.size() - 1);
+
+            for (uint32_t i = firstIndex; i <= lastIndex && filterBankIter != end; ++i) {
+                float energyRep = math::MathUtils::SqrtF32(fftVec[i]);
+                melEnergy += (*filterBankIter++ * energyRep);
+            }
+
+            melEnergies[bin] = melEnergy;
+        }
+
+        return true;
+    }
+
+    void MelSpectrogram::ConvertToLogarithmicScale(std::vector<float>& melEnergies)
+    {
+        for (float& melEnergy : melEnergies) {
+            melEnergy = logf(melEnergy);
+        }
+    }
+
+    void MelSpectrogram::ConvertToPowerSpectrum()
+    {
+        const uint32_t halfDim = this->m_buffer.size() / 2;
+
+        /* Handle this special case. */
+        float firstEnergy = this->m_buffer[0] * this->m_buffer[0];
+        float lastEnergy = this->m_buffer[1] * this->m_buffer[1];
+
+        math::MathUtils::ComplexMagnitudeSquaredF32(
+                this->m_buffer.data(),
+                this->m_buffer.size(),
+                this->m_buffer.data(),
+                this->m_buffer.size()/2);
+
+        this->m_buffer[0] = firstEnergy;
+        this->m_buffer[halfDim] = lastEnergy;
+    }
+
+    float MelSpectrogram::GetMelFilterBankNormaliser(
+            const float&    leftMel,
+            const float&    rightMel,
+            const bool      useHTKMethod)
+    {
+        UNUSED(leftMel);
+        UNUSED(rightMel);
+        UNUSED(useHTKMethod);
+
+        /* By default, no normalisation => return 1 */
+        return 1.f;
+    }
+
+    void MelSpectrogram::InitMelFilterBank()
+    {
+        if (!this->IsMelFilterBankInited()) {
+            this->m_melFilterBank = this->CreateMelFilterBank();
+            this->m_filterBankInitialised = true;
+        }
+    }
+
+    bool MelSpectrogram::IsMelFilterBankInited() const
+    {
+        return this->m_filterBankInitialised;
+    }
+
+    std::vector<float> MelSpectrogram::ComputeMelSpec(const std::vector<int16_t>& audioData, float trainingMean)
+    {
+        this->InitMelFilterBank();
+
+        /* TensorFlow way of normalizing .wav data to (-1, 1). */
+        constexpr float normaliser = 1.0/(1<<15);
+        for (size_t i = 0; i < this->m_params.m_frameLen; ++i) {
+            this->m_frame[i] = static_cast<float>(audioData[i]) * normaliser;
+        }
+
+        /* Apply window function to input frame. */
+        for(size_t i = 0; i < this->m_params.m_frameLen; ++i) {
+            this->m_frame[i] *= this->m_windowFunc[i];
+        }
+
+        /* Set remaining frame values to 0. */
+        std::fill(this->m_frame.begin() + this->m_params.m_frameLen,this->m_frame.end(), 0);
+
+        /* Compute FFT. */
+        math::MathUtils::FftF32(this->m_frame, this->m_buffer, this->m_fftInstance);
+
+        /* Convert to power spectrum. */
+        this->ConvertToPowerSpectrum();
+
+        /* Apply mel filterbanks. */
+        if (!this->ApplyMelFilterBank(this->m_buffer,
+                                      this->m_melFilterBank,
+                                      this->m_filterBankFilterFirst,
+                                      this->m_filterBankFilterLast,
+                                      this->m_melEnergies)) {
+            printf_err("Failed to apply MEL filter banks\n");
+        }
+
+        /* Convert to logarithmic scale */
+        this->ConvertToLogarithmicScale(this->m_melEnergies);
+
+        /* Perform mean subtraction. */
+        for (auto& energy:this->m_melEnergies) {
+            energy -= trainingMean;
+        }
+
+        return this->m_melEnergies;
+    }
+
+    std::vector<std::vector<float>> MelSpectrogram::CreateMelFilterBank()
+    {
+        size_t numFftBins = this->m_params.m_frameLenPadded / 2;
+        float fftBinWidth = static_cast<float>(this->m_params.m_samplingFreq) / this->m_params.m_frameLenPadded;
+
+        float melLowFreq = MelSpectrogram::MelScale(this->m_params.m_melLoFreq,
+                                          this->m_params.m_useHtkMethod);
+        float melHighFreq = MelSpectrogram::MelScale(this->m_params.m_melHiFreq,
+                                           this->m_params.m_useHtkMethod);
+        float melFreqDelta = (melHighFreq - melLowFreq) / (this->m_params.m_numFbankBins + 1);
+
+        std::vector<float> thisBin = std::vector<float>(numFftBins);
+        std::vector<std::vector<float>> melFilterBank(
+                this->m_params.m_numFbankBins);
+        this->m_filterBankFilterFirst =
+                std::vector<uint32_t>(this->m_params.m_numFbankBins);
+        this->m_filterBankFilterLast =
+                std::vector<uint32_t>(this->m_params.m_numFbankBins);
+
+        for (size_t bin = 0; bin < this->m_params.m_numFbankBins; bin++) {
+            float leftMel = melLowFreq + bin * melFreqDelta;
+            float centerMel = melLowFreq + (bin + 1) * melFreqDelta;
+            float rightMel = melLowFreq + (bin + 2) * melFreqDelta;
+
+            uint32_t firstIndex = 0;
+            uint32_t lastIndex = 0;
+            bool firstIndexFound = false;
+            const float normaliser = this->GetMelFilterBankNormaliser(leftMel, rightMel, this->m_params.m_useHtkMethod);
+
+            for (size_t i = 0; i < numFftBins; ++i) {
+                float freq = (fftBinWidth * i); /* Center freq of this fft bin. */
+                float mel = MelSpectrogram::MelScale(freq, this->m_params.m_useHtkMethod);
+                thisBin[i] = 0.0;
+
+                if (mel > leftMel && mel < rightMel) {
+                    float weight;
+                    if (mel <= centerMel) {
+                        weight = (mel - leftMel) / (centerMel - leftMel);
+                    } else {
+                        weight = (rightMel - mel) / (rightMel - centerMel);
+                    }
+
+                    thisBin[i] = weight * normaliser;
+                    if (!firstIndexFound) {
+                        firstIndex = i;
+                        firstIndexFound = true;
+                    }
+                    lastIndex = i;
+                }
+            }
+
+            this->m_filterBankFilterFirst[bin] = firstIndex;
+            this->m_filterBankFilterLast[bin] = lastIndex;
+
+            /* Copy the part we care about. */
+            for (uint32_t i = firstIndex; i <= lastIndex; ++i) {
+                melFilterBank[bin].push_back(thisBin[i]);
+            }
+        }
+
+        return melFilterBank;
+    }
+
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
diff --git a/source/application/api/use_case/asr/CMakeLists.txt b/source/application/api/use_case/asr/CMakeLists.txt
new file mode 100644
index 0000000..77e3d6a
--- /dev/null
+++ b/source/application/api/use_case/asr/CMakeLists.txt
@@ -0,0 +1,43 @@
+#----------------------------------------------------------------------------
+#  Copyright (c) 2022 Arm Limited. All rights reserved.
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#----------------------------------------------------------------------------
+#########################################################
+#       AUTOMATIC SPEECH RECOGNITION API library        #
+#########################################################
+cmake_minimum_required(VERSION 3.15.6)
+
+set(ASR_API_TARGET asr_api)
+project(${ASR_API_TARGET}
+        DESCRIPTION     "Automatic speech recognition use case API library"
+        LANGUAGES       C CXX)
+
+# Create static library
+add_library(${ASR_API_TARGET} STATIC
+        src/Wav2LetterPreprocess.cc
+        src/Wav2LetterPostprocess.cc
+        src/Wav2LetterMfcc.cc
+        src/AsrClassifier.cc
+        src/OutputDecode.cc
+        src/Wav2LetterModel.cc)
+
+target_include_directories(${ASR_API_TARGET} PUBLIC include)
+
+target_link_libraries(${ASR_API_TARGET} PUBLIC common_api)
+
+message(STATUS "*******************************************************")
+message(STATUS "Library                                : " ${ASR_API_TARGET})
+message(STATUS "CMAKE_SYSTEM_PROCESSOR                 : " ${CMAKE_SYSTEM_PROCESSOR})
+message(STATUS "*******************************************************")
diff --git a/source/application/api/use_case/asr/include/AsrClassifier.hpp b/source/application/api/use_case/asr/include/AsrClassifier.hpp
new file mode 100644
index 0000000..a07a721
--- /dev/null
+++ b/source/application/api/use_case/asr/include/AsrClassifier.hpp
@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef ASR_CLASSIFIER_HPP
+#define ASR_CLASSIFIER_HPP
+
+#include "Classifier.hpp"
+
+namespace arm {
+namespace app {
+
+    class AsrClassifier : public Classifier {
+    public:
+        /**
+         * @brief       Gets the top N classification results from the
+         *              output vector.
+         * @param[in]   outputTensor   Inference output tensor from an NN model.
+         * @param[out]  vecResults     A vector of classification results
+         *                             populated by this function.
+         * @param[in]   labels         Labels vector to match classified classes
+         * @param[in]   topNCount      Number of top classifications to pick.
+         * @param[in]   use_softmax    Whether softmax scaling should be applied to model output.
+         * @return      true if successful, false otherwise.
+         **/
+        bool GetClassificationResults(TfLiteTensor* outputTensor,
+                                      std::vector<ClassificationResult>& vecResults,
+                                      const std::vector<std::string>& labels,
+                                      uint32_t topNCount, bool use_softmax = false) override;
+
+    private:
+        /**
+         * @brief       Utility function that gets the top 1 classification results from the
+         *              output tensor (vector of vector).
+         * @param[in]   tensor       Inference output tensor from an NN model.
+         * @param[out]  vecResults   Vector of classification results populated by this function.
+         * @param[in]   labels       Labels vector to match classified classes.
+         * @param[in]   scale        Quantization scale.
+         * @param[in]   zeroPoint    Quantization zero point.
+         * @return      true if successful, false otherwise.
+         **/
+        template<typename T>
+        bool GetTopResults(TfLiteTensor* tensor,
+                           std::vector<ClassificationResult>& vecResults,
+                           const std::vector<std::string>& labels, double scale, double zeroPoint);
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* ASR_CLASSIFIER_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/asr/include/AsrResult.hpp b/source/application/api/use_case/asr/include/AsrResult.hpp
new file mode 100644
index 0000000..ed826d0
--- /dev/null
+++ b/source/application/api/use_case/asr/include/AsrResult.hpp
@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef ASR_RESULT_HPP
+#define ASR_RESULT_HPP
+
+#include "ClassificationResult.hpp"
+
+#include <vector>
+
+namespace arm {
+namespace app {
+namespace asr {
+
+    using ResultVec = std::vector<arm::app::ClassificationResult>;
+
+    /* Structure for holding ASR result. */
+    class AsrResult {
+
+    public:
+        ResultVec       m_resultVec;        /* Container for "thresholded" classification results. */
+        float           m_timeStamp;        /* Audio timestamp for this result. */
+        uint32_t        m_inferenceNumber;  /* Corresponding inference number. */
+        float           m_threshold;        /* Threshold value for `m_resultVec.` */
+
+        AsrResult() = delete;
+        AsrResult(ResultVec&        resultVec,
+                  const float       timestamp,
+                  const uint32_t    inferenceIdx,
+                  const float       scoreThreshold) {
+
+            this->m_threshold = scoreThreshold;
+            this->m_timeStamp = timestamp;
+            this->m_inferenceNumber = inferenceIdx;
+
+            this->m_resultVec = ResultVec();
+            for (auto& i : resultVec) {
+                if (i.m_normalisedVal >= this->m_threshold) {
+                    this->m_resultVec.emplace_back(i);
+                }
+            }
+        }
+        ~AsrResult() = default;
+    };
+
+} /* namespace asr */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* ASR_RESULT_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/asr/include/OutputDecode.hpp b/source/application/api/use_case/asr/include/OutputDecode.hpp
new file mode 100644
index 0000000..9d39057
--- /dev/null
+++ b/source/application/api/use_case/asr/include/OutputDecode.hpp
@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef ASR_OUTPUT_DECODE_HPP
+#define ASR_OUTPUT_DECODE_HPP
+
+#include "AsrClassifier.hpp"
+
+namespace arm {
+namespace app {
+namespace audio {
+namespace asr {
+
+    /**
+     * @brief       Gets the top N classification results from the
+     *              output vector.
+     * @param[in]   vecResults   Label output from classifier.
+     * @return      true if successful, false otherwise.
+    **/
+    std::string DecodeOutput(const std::vector<ClassificationResult>& vecResults);
+
+} /* namespace asr */
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* ASR_OUTPUT_DECODE_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/asr/include/Wav2LetterMfcc.hpp b/source/application/api/use_case/asr/include/Wav2LetterMfcc.hpp
new file mode 100644
index 0000000..b5a21d3
--- /dev/null
+++ b/source/application/api/use_case/asr/include/Wav2LetterMfcc.hpp
@@ -0,0 +1,109 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef ASR_WAV2LETTER_MFCC_HPP
+#define ASR_WAV2LETTER_MFCC_HPP
+
+#include "Mfcc.hpp"
+
+namespace arm {
+namespace app {
+namespace audio {
+
+    /* Class to provide Wav2Letter specific MFCC calculation requirements. */
+    class Wav2LetterMFCC : public MFCC {
+
+    public:
+        static constexpr uint32_t  ms_defaultSamplingFreq = 16000;
+        static constexpr uint32_t  ms_defaultNumFbankBins =   128;
+        static constexpr uint32_t  ms_defaultMelLoFreq    =     0;
+        static constexpr uint32_t  ms_defaultMelHiFreq    =  8000;
+        static constexpr bool      ms_defaultUseHtkMethod = false;
+
+        explicit Wav2LetterMFCC(const size_t numFeats, const size_t frameLen)
+            :  MFCC(MfccParams(
+                        ms_defaultSamplingFreq, ms_defaultNumFbankBins,
+                        ms_defaultMelLoFreq, ms_defaultMelHiFreq,
+                        numFeats, frameLen, ms_defaultUseHtkMethod))
+        {}
+
+        Wav2LetterMFCC()  = delete;
+        ~Wav2LetterMFCC() = default;
+
+    protected:
+
+        /**
+         * @brief       Overrides base class implementation of this function.
+         * @param[in]   fftVec                  Vector populated with FFT magnitudes
+         * @param[in]   melFilterBank           2D Vector with filter bank weights
+         * @param[in]   filterBankFilterFirst   Vector containing the first indices of filter bank
+         *                                      to be used for each bin.
+         * @param[in]   filterBankFilterLast    Vector containing the last indices of filter bank
+         *                                      to be used for each bin.
+         * @param[out]  melEnergies             Pre-allocated vector of MEL energies to be
+         *                                      populated.
+         * @return      true if successful, false otherwise
+         */
+        bool ApplyMelFilterBank(
+            std::vector<float>&                 fftVec,
+            std::vector<std::vector<float>>&    melFilterBank,
+            std::vector<uint32_t>&              filterBankFilterFirst,
+            std::vector<uint32_t>&              filterBankFilterLast,
+            std::vector<float>&                 melEnergies) override;
+
+        /**
+         * @brief           Override for the base class implementation convert mel
+         *                  energies to logarithmic scale. The difference from
+         *                  default behaviour is that the power is converted to dB
+         *                  and subsequently clamped.
+         * @param[in,out]   melEnergies   1D vector of Mel energies
+         **/
+        void ConvertToLogarithmicScale(std::vector<float>& melEnergies) override;
+
+        /**
+         * @brief       Create a matrix used to calculate Discrete Cosine
+         *              Transform. Override for the base class' default
+         *              implementation as the first and last elements
+         *              use a different normaliser.
+         * @param[in]   inputLength        input length of the buffer on which
+         *                                 DCT will be performed
+         * @param[in]   coefficientCount   Total coefficients per input length.
+         * @return      1D vector with inputLength x coefficientCount elements
+         *              populated with DCT coefficients.
+         */
+        std::vector<float> CreateDCTMatrix(int32_t inputLength,
+                                           int32_t coefficientCount) override;
+
+        /**
+         * @brief       Given the low and high Mel values, get the normaliser
+         *              for weights to be applied when populating the filter
+         *              bank. Override for the base class implementation.
+         * @param[in]   leftMel        Low Mel frequency value.
+         * @param[in]   rightMel       High Mel frequency value.
+         * @param[in]   useHTKMethod   bool to signal if HTK method is to be
+         *                             used for calculation.
+         * @return      Value to use for normalising.
+         */
+        float GetMelFilterBankNormaliser(const float&   leftMel,
+                                         const float&   rightMel,
+                                         bool     useHTKMethod) override;
+    };
+
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* ASR_WAV2LETTER_MFCC_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/asr/include/Wav2LetterModel.hpp b/source/application/api/use_case/asr/include/Wav2LetterModel.hpp
new file mode 100644
index 0000000..a02eed1
--- /dev/null
+++ b/source/application/api/use_case/asr/include/Wav2LetterModel.hpp
@@ -0,0 +1,67 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef ASR_WAV2LETTER_MODEL_HPP
+#define ASR_WAV2LETTER_MODEL_HPP
+
+#include "Model.hpp"
+
+namespace arm {
+namespace app {
+namespace asr {
+    extern const int g_FrameLength;
+    extern const int g_FrameStride;
+    extern const float g_ScoreThreshold;
+    extern const int g_ctxLen;
+} /* namespace asr */
+} /* namespace app */
+} /* namespace arm */
+
+namespace arm {
+namespace app {
+
+    class Wav2LetterModel : public Model {
+
+    public:
+        /* Indices for the expected model - based on input and output tensor shapes */
+        static constexpr uint32_t ms_inputRowsIdx  = 1;
+        static constexpr uint32_t ms_inputColsIdx  = 2;
+        static constexpr uint32_t ms_outputRowsIdx = 2;
+        static constexpr uint32_t ms_outputColsIdx = 3;
+
+        /* Model specific constants. */
+        static constexpr uint32_t ms_blankTokenIdx   = 28;
+        static constexpr uint32_t ms_numMfccFeatures = 13;
+
+    protected:
+        /** @brief   Gets the reference to op resolver interface class. */
+        const tflite::MicroOpResolver& GetOpResolver() override;
+
+        /** @brief   Adds operations to the op resolver instance. */
+        bool EnlistOperations() override;
+
+    private:
+        /* Maximum number of individual operations that can be enlisted. */
+        static constexpr int ms_maxOpCnt = 5;
+
+        /* A mutable op resolver instance. */
+        tflite::MicroMutableOpResolver<ms_maxOpCnt> m_opResolver;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* ASR_WAV2LETTER_MODEL_HPP */
diff --git a/source/application/api/use_case/asr/include/Wav2LetterPostprocess.hpp b/source/application/api/use_case/asr/include/Wav2LetterPostprocess.hpp
new file mode 100644
index 0000000..02738bc
--- /dev/null
+++ b/source/application/api/use_case/asr/include/Wav2LetterPostprocess.hpp
@@ -0,0 +1,109 @@
+/*
+ * Copyright (c) 2021-2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef ASR_WAV2LETTER_POSTPROCESS_HPP
+#define ASR_WAV2LETTER_POSTPROCESS_HPP
+
+#include "TensorFlowLiteMicro.hpp"   /* TensorFlow headers. */
+#include "BaseProcessing.hpp"
+#include "Model.hpp"
+#include "AsrClassifier.hpp"
+#include "AsrResult.hpp"
+#include "log_macros.h"
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   Helper class to manage tensor post-processing for "wav2letter"
+     *          output.
+     */
+    class AsrPostProcess : public BasePostProcess {
+    public:
+        bool m_lastIteration = false;   /* Flag to set if processing the last set of data for a clip. */
+
+        /**
+         * @brief           Constructor
+         * @param[in]       outputTensor       Pointer to the TFLite Micro output Tensor.
+         * @param[in]       classifier         Object used to get top N results from classification.
+         * @param[in]       labels             Vector of string labels to identify each output of the model.
+         * @param[in/out]   result             Vector of classification results to store decoded outputs.
+         * @param[in]       outputContextLen   Left/right context length for output tensor.
+         * @param[in]       blankTokenIdx      Index in the labels that the "Blank token" takes.
+         * @param[in]       reductionAxis      The axis that the logits of each time step is on.
+         **/
+        AsrPostProcess(TfLiteTensor* outputTensor, AsrClassifier& classifier,
+                       const std::vector<std::string>& labels, asr::ResultVec& result,
+                       uint32_t outputContextLen,
+                       uint32_t blankTokenIdx, uint32_t reductionAxis);
+
+        /**
+         * @brief    Should perform post-processing of the result of inference then
+         *           populate ASR result data for any later use.
+         * @return   true if successful, false otherwise.
+         **/
+        bool DoPostProcess() override;
+
+        /** @brief   Gets the output inner length for post-processing. */
+        static uint32_t GetOutputInnerLen(const TfLiteTensor*, uint32_t outputCtxLen);
+
+        /** @brief   Gets the output context length (left/right) for post-processing. */
+        static uint32_t GetOutputContextLen(const Model& model, uint32_t inputCtxLen);
+
+        /** @brief   Gets the number of feature vectors to be computed. */
+        static uint32_t GetNumFeatureVectors(const Model& model);
+
+    private:
+        AsrClassifier& m_classifier;                /* ASR Classifier object. */
+        TfLiteTensor* m_outputTensor;               /* Model output tensor. */
+        const std::vector<std::string>& m_labels;   /* ASR Labels. */
+        asr::ResultVec & m_results;                 /* Results vector for a single inference. */
+        uint32_t m_outputContextLen;                /* lengths of left/right contexts for output. */
+        uint32_t m_outputInnerLen;                  /* Length of output inner context. */
+        uint32_t m_totalLen;                        /* Total length of the required axis. */
+        uint32_t m_countIterations;                 /* Current number of iterations. */
+        uint32_t m_blankTokenIdx;                   /* Index of the labels blank token. */
+        uint32_t m_reductionAxisIdx;                /* Axis containing output logits for a single step. */
+
+        /**
+         * @brief    Checks if the tensor and axis index are valid
+         *           inputs to the object - based on how it has been initialised.
+         * @return   true if valid, false otherwise.
+         */
+        bool IsInputValid(TfLiteTensor*  tensor,
+                          uint32_t axisIdx) const;
+
+        /**
+         * @brief    Gets the tensor data element size in bytes based
+         *           on the tensor type.
+         * @return   Size in bytes, 0 if not supported.
+         */
+        static uint32_t GetTensorElementSize(TfLiteTensor* tensor);
+
+        /**
+         * @brief    Erases sections from the data assuming row-wise
+         *           arrangement along the context axis.
+         * @return   true if successful, false otherwise.
+         */
+        bool EraseSectionsRowWise(uint8_t* ptrData,
+                                  uint32_t strideSzBytes,
+                                  bool lastIteration);
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* ASR_WAV2LETTER_POSTPROCESS_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/asr/include/Wav2LetterPreprocess.hpp b/source/application/api/use_case/asr/include/Wav2LetterPreprocess.hpp
new file mode 100644
index 0000000..9943946
--- /dev/null
+++ b/source/application/api/use_case/asr/include/Wav2LetterPreprocess.hpp
@@ -0,0 +1,182 @@
+/*
+ * Copyright (c) 2021-2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef ASR_WAV2LETTER_PREPROCESS_HPP
+#define ASR_WAV2LETTER_PREPROCESS_HPP
+
+#include "TensorFlowLiteMicro.hpp"
+#include "Wav2LetterMfcc.hpp"
+#include "AudioUtils.hpp"
+#include "DataStructures.hpp"
+#include "BaseProcessing.hpp"
+#include "log_macros.h"
+
+namespace arm {
+namespace app {
+
+    /* Class to facilitate pre-processing calculation for Wav2Letter model
+     * for ASR. */
+    using AudioWindow = audio::SlidingWindow<const int16_t>;
+
+    class AsrPreProcess : public BasePreProcess {
+    public:
+        /**
+         * @brief       Constructor.
+         * @param[in]   inputTensor        Pointer to the TFLite Micro input Tensor.
+         * @param[in]   numMfccFeatures    Number of MFCC features per window.
+         * @param[in]   numFeatureFrames   Number of MFCC vectors that need to be calculated
+         *                                 for an inference.
+         * @param[in]   mfccWindowLen      Number of audio elements to calculate MFCC features per window.
+         * @param[in]   mfccWindowStride   Stride (in number of elements) for moving the MFCC window.
+         */
+        AsrPreProcess(TfLiteTensor* inputTensor,
+                      uint32_t  numMfccFeatures,
+                      uint32_t  numFeatureFrames,
+                      uint32_t  mfccWindowLen,
+                      uint32_t  mfccWindowStride);
+
+        /**
+         * @brief       Calculates the features required from audio data. This
+         *              includes MFCC, first and second order deltas,
+         *              normalisation and finally, quantisation. The tensor is
+         *              populated with features from a given window placed along
+         *              in a single row.
+         * @param[in]   audioData      Pointer to the first element of audio data.
+         * @param[in]   audioDataLen   Number of elements in the audio data.
+         * @return      true if successful, false in case of error.
+         */
+        bool DoPreProcess(const void* audioData, size_t audioDataLen) override;
+
+    protected:
+         /**
+          * @brief Computes the first and second order deltas for the
+          *        MFCC buffers - they are assumed to be populated.
+          *
+          * @param[in]  mfcc     MFCC buffers.
+          * @param[out] delta1   Result of the first diff computation.
+          * @param[out] delta2   Result of the second diff computation.
+          * @return     true if successful, false otherwise.
+          */
+         static bool ComputeDeltas(Array2d<float>& mfcc,
+                                   Array2d<float>& delta1,
+                                   Array2d<float>& delta2);
+
+        /**
+         * @brief           Given a 2D vector of floats, rescale it to have mean of 0 and
+        *                   standard deviation of 1.
+         * @param[in,out]   vec   Vector of vector of floats.
+         */
+        static void StandardizeVecF32(Array2d<float>& vec);
+
+        /**
+         * @brief   Standardizes all the MFCC and delta buffers to have mean 0 and std. dev 1.
+         */
+        void Standarize();
+
+        /**
+         * @brief       Given the quantisation and data type limits, computes
+         *              the quantised values of a floating point input data.
+         * @param[in]   elem          Element to be quantised.
+         * @param[in]   quantScale    Scale.
+         * @param[in]   quantOffset   Offset.
+         * @param[in]   minVal        Numerical limit - minimum.
+         * @param[in]   maxVal        Numerical limit - maximum.
+         * @return      Floating point quantised value.
+         */
+        static float GetQuantElem(
+                float     elem,
+                float     quantScale,
+                int       quantOffset,
+                float     minVal,
+                float     maxVal);
+
+        /**
+         * @brief       Quantises the MFCC and delta buffers, and places them
+         *              in the output buffer. While doing so, it transposes
+         *              the data. Reason: Buffers in this class are arranged
+         *              for "time" axis to be row major. Primary reason for
+         *              this being the convolution speed up (as we can use
+         *              contiguous memory). The output, however, requires the
+         *              time axis to be in column major arrangement.
+         * @param[in]   outputBuf     Pointer to the output buffer.
+         * @param[in]   outputBufSz   Output buffer's size.
+         * @param[in]   quantScale    Quantisation scale.
+         * @param[in]   quantOffset   Quantisation offset.
+         */
+        template <typename T>
+        bool Quantise(
+                T*              outputBuf,
+                const uint32_t  outputBufSz,
+                const float     quantScale,
+                const int       quantOffset)
+        {
+            /* Check the output size will fit everything. */
+            if (outputBufSz < (this->m_mfccBuf.size(0) * 3 * sizeof(T))) {
+                printf_err("Tensor size too small for features\n");
+                return false;
+            }
+
+            /* Populate. */
+            T* outputBufMfcc = outputBuf;
+            T* outputBufD1 = outputBuf + this->m_numMfccFeats;
+            T* outputBufD2 = outputBufD1 + this->m_numMfccFeats;
+            const uint32_t ptrIncr = this->m_numMfccFeats * 2;  /* (3 vectors - 1 vector) */
+
+            const float minVal = std::numeric_limits<T>::min();
+            const float maxVal = std::numeric_limits<T>::max();
+
+            /* Need to transpose while copying and concatenating the tensor. */
+            for (uint32_t j = 0; j < this->m_numFeatureFrames; ++j) {
+                for (uint32_t i = 0; i < this->m_numMfccFeats; ++i) {
+                    *outputBufMfcc++ = static_cast<T>(AsrPreProcess::GetQuantElem(
+                            this->m_mfccBuf(i, j), quantScale,
+                            quantOffset, minVal, maxVal));
+                    *outputBufD1++ = static_cast<T>(AsrPreProcess::GetQuantElem(
+                            this->m_delta1Buf(i, j), quantScale,
+                            quantOffset, minVal, maxVal));
+                    *outputBufD2++ = static_cast<T>(AsrPreProcess::GetQuantElem(
+                            this->m_delta2Buf(i, j), quantScale,
+                            quantOffset, minVal, maxVal));
+                }
+                outputBufMfcc += ptrIncr;
+                outputBufD1 += ptrIncr;
+                outputBufD2 += ptrIncr;
+            }
+
+            return true;
+        }
+
+    private:
+        audio::Wav2LetterMFCC   m_mfcc;          /* MFCC instance. */
+        TfLiteTensor*           m_inputTensor;   /* Model input tensor. */
+
+        /* Actual buffers to be populated. */
+        Array2d<float>   m_mfccBuf;              /* Contiguous buffer 1D: MFCC */
+        Array2d<float>   m_delta1Buf;            /* Contiguous buffer 1D: Delta 1 */
+        Array2d<float>   m_delta2Buf;            /* Contiguous buffer 1D: Delta 2 */
+
+        uint32_t         m_mfccWindowLen;        /* Window length for MFCC. */
+        uint32_t         m_mfccWindowStride;     /* Window stride len for MFCC. */
+        uint32_t         m_numMfccFeats;         /* Number of MFCC features per window. */
+        uint32_t         m_numFeatureFrames;     /* How many sets of m_numMfccFeats. */
+        AudioWindow      m_mfccSlidingWindow;    /* Sliding window to calculate MFCCs. */
+
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* ASR_WAV2LETTER_PREPROCESS_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/asr/src/AsrClassifier.cc b/source/application/api/use_case/asr/src/AsrClassifier.cc
new file mode 100644
index 0000000..4ba8c7b
--- /dev/null
+++ b/source/application/api/use_case/asr/src/AsrClassifier.cc
@@ -0,0 +1,144 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "AsrClassifier.hpp"
+
+#include "log_macros.h"
+#include "TensorFlowLiteMicro.hpp"
+#include "Wav2LetterModel.hpp"
+
+namespace arm {
+namespace app {
+
+    template<typename T>
+    bool AsrClassifier::GetTopResults(TfLiteTensor* tensor,
+                                      std::vector<ClassificationResult>& vecResults,
+                                      const std::vector <std::string>& labels, double scale, double zeroPoint)
+    {
+        const uint32_t nElems = tensor->dims->data[Wav2LetterModel::ms_outputRowsIdx];
+        const uint32_t nLetters = tensor->dims->data[Wav2LetterModel::ms_outputColsIdx];
+
+        if (nLetters != labels.size()) {
+            printf("Output size doesn't match the labels' size\n");
+            return false;
+        }
+
+        /* NOTE: tensor's size verification against labels should be
+         *       checked by the calling/public function. */
+        if (nLetters < 1) {
+            return false;
+        }
+
+        /* Final results' container. */
+        vecResults = std::vector<ClassificationResult>(nElems);
+
+        T* tensorData = tflite::GetTensorData<T>(tensor);
+
+        /* Get the top 1 results. */
+        for (uint32_t i = 0, row = 0; i < nElems; ++i, row+=nLetters) {
+            std::pair<T, uint32_t> top_1 = std::make_pair(tensorData[row + 0], 0);
+
+            for (uint32_t j = 1; j < nLetters; ++j) {
+                if (top_1.first < tensorData[row + j]) {
+                    top_1.first = tensorData[row + j];
+                    top_1.second = j;
+                }
+            }
+
+            double score = static_cast<int> (top_1.first);
+            vecResults[i].m_normalisedVal = scale * (score - zeroPoint);
+            vecResults[i].m_label = labels[top_1.second];
+            vecResults[i].m_labelIdx = top_1.second;
+        }
+
+        return true;
+    }
+    template bool AsrClassifier::GetTopResults<uint8_t>(TfLiteTensor* tensor,
+                                                        std::vector<ClassificationResult>& vecResults,
+                                                        const std::vector <std::string>& labels,
+                                                        double scale, double zeroPoint);
+    template bool AsrClassifier::GetTopResults<int8_t>(TfLiteTensor* tensor,
+                                                       std::vector<ClassificationResult>& vecResults,
+                                                       const std::vector <std::string>& labels,
+                                                       double scale, double zeroPoint);
+
+    bool AsrClassifier::GetClassificationResults(
+            TfLiteTensor* outputTensor,
+            std::vector<ClassificationResult>& vecResults,
+            const std::vector <std::string>& labels, uint32_t topNCount, bool use_softmax)
+    {
+            UNUSED(use_softmax);
+            vecResults.clear();
+
+            constexpr int minTensorDims = static_cast<int>(
+                (Wav2LetterModel::ms_outputRowsIdx > Wav2LetterModel::ms_outputColsIdx)?
+                 Wav2LetterModel::ms_outputRowsIdx : Wav2LetterModel::ms_outputColsIdx);
+
+            constexpr uint32_t outColsIdx = Wav2LetterModel::ms_outputColsIdx;
+
+            /* Sanity checks. */
+            if (outputTensor == nullptr) {
+                printf_err("Output vector is null pointer.\n");
+                return false;
+            } else if (outputTensor->dims->size < minTensorDims) {
+                printf_err("Output tensor expected to be %dD\n", minTensorDims);
+                return false;
+            } else if (static_cast<uint32_t>(outputTensor->dims->data[outColsIdx]) < topNCount) {
+                printf_err("Output vectors are smaller than %" PRIu32 "\n", topNCount);
+                return false;
+            } else if (static_cast<uint32_t>(outputTensor->dims->data[outColsIdx]) != labels.size()) {
+                printf("Output size doesn't match the labels' size\n");
+                return false;
+            }
+
+            if (topNCount != 1) {
+                warn("TopNCount value ignored in this implementation\n");
+            }
+
+            /* To return the floating point values, we need quantization parameters. */
+            QuantParams quantParams = GetTensorQuantParams(outputTensor);
+
+            bool resultState;
+
+            switch (outputTensor->type) {
+                case kTfLiteUInt8:
+                    resultState = this->GetTopResults<uint8_t>(
+                            outputTensor, vecResults,
+                            labels, quantParams.scale,
+                            quantParams.offset);
+                    break;
+                case kTfLiteInt8:
+                    resultState = this->GetTopResults<int8_t>(
+                            outputTensor, vecResults,
+                            labels, quantParams.scale,
+                            quantParams.offset);
+                    break;
+                default:
+                    printf_err("Tensor type %s not supported by classifier\n",
+                        TfLiteTypeGetName(outputTensor->type));
+                    return false;
+            }
+
+            if (!resultState) {
+                printf_err("Failed to get sorted set\n");
+                return false;
+            }
+
+            return true;
+    }
+
+} /* namespace app */
+} /* namespace arm */
\ No newline at end of file
diff --git a/source/application/api/use_case/asr/src/OutputDecode.cc b/source/application/api/use_case/asr/src/OutputDecode.cc
new file mode 100644
index 0000000..41fbe07
--- /dev/null
+++ b/source/application/api/use_case/asr/src/OutputDecode.cc
@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "OutputDecode.hpp"
+
+namespace arm {
+namespace app {
+namespace audio {
+namespace asr {
+
+    std::string DecodeOutput(const std::vector<ClassificationResult>& vecResults)
+    {
+        std::string CleanOutputBuffer;
+
+        for (size_t i = 0; i < vecResults.size(); ++i)  /* For all elements in vector. */
+        {
+            while (i+1 < vecResults.size() &&
+                   vecResults[i].m_label == vecResults[i+1].m_label)  /* While the current element is equal to the next, ignore it and move on. */
+            {
+                ++i;
+            }
+            if (vecResults[i].m_label != "$")  /* $ is a character used to represent unknown and double characters so should not be in output. */
+            {
+                CleanOutputBuffer += vecResults[i].m_label;  /* If the element is different to the next, it will be appended to CleanOutputBuffer. */
+            }
+        }
+
+        return CleanOutputBuffer;  /* Return string type containing clean output. */
+    }
+
+} /* namespace asr */
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
diff --git a/source/application/api/use_case/asr/src/Wav2LetterMfcc.cc b/source/application/api/use_case/asr/src/Wav2LetterMfcc.cc
new file mode 100644
index 0000000..bb29b0f
--- /dev/null
+++ b/source/application/api/use_case/asr/src/Wav2LetterMfcc.cc
@@ -0,0 +1,141 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "Wav2LetterMfcc.hpp"
+
+#include "PlatformMath.hpp"
+#include "log_macros.h"
+
+#include <cfloat>
+
+namespace arm {
+namespace app {
+namespace audio {
+
+    bool Wav2LetterMFCC::ApplyMelFilterBank(
+            std::vector<float>&                 fftVec,
+            std::vector<std::vector<float>>&    melFilterBank,
+            std::vector<uint32_t>&               filterBankFilterFirst,
+            std::vector<uint32_t>&               filterBankFilterLast,
+            std::vector<float>&                 melEnergies)
+    {
+        const size_t numBanks = melEnergies.size();
+
+        if (numBanks != filterBankFilterFirst.size() ||
+                numBanks != filterBankFilterLast.size()) {
+            printf_err("Unexpected filter bank lengths\n");
+            return false;
+        }
+
+        for (size_t bin = 0; bin < numBanks; ++bin) {
+            auto filterBankIter = melFilterBank[bin].begin();
+            auto end = melFilterBank[bin].end();
+            /* Avoid log of zero at later stages, same value used in librosa.
+             * The number was used during our default wav2letter model training. */
+            float melEnergy = 1e-10;
+            const uint32_t firstIndex = filterBankFilterFirst[bin];
+            const uint32_t lastIndex = std::min<uint32_t>(filterBankFilterLast[bin], fftVec.size() - 1);
+
+            for (uint32_t i = firstIndex; i <= lastIndex && filterBankIter != end; ++i) {
+                melEnergy += (*filterBankIter++ * fftVec[i]);
+            }
+
+            melEnergies[bin] = melEnergy;
+        }
+
+        return true;
+    }
+
+    void Wav2LetterMFCC::ConvertToLogarithmicScale(
+                            std::vector<float>& melEnergies)
+    {
+        float maxMelEnergy = -FLT_MAX;
+
+        /* Container for natural logarithms of mel energies. */
+        std::vector <float> vecLogEnergies(melEnergies.size(), 0.f);
+
+        /* Because we are taking natural logs, we need to multiply by log10(e).
+         * Also, for wav2letter model, we scale our log10 values by 10. */
+        constexpr float multiplier = 10.0 *  /* Default scalar. */
+                                      0.4342944819032518;  /* log10f(std::exp(1.0)) */
+
+        /* Take log of the whole vector. */
+        math::MathUtils::VecLogarithmF32(melEnergies, vecLogEnergies);
+
+        /* Scale the log values and get the max. */
+        for (auto iterM = melEnergies.begin(), iterL = vecLogEnergies.begin();
+                  iterM != melEnergies.end() && iterL != vecLogEnergies.end(); ++iterM, ++iterL) {
+
+            *iterM = *iterL * multiplier;
+
+            /* Save the max mel energy. */
+            if (*iterM > maxMelEnergy) {
+                maxMelEnergy = *iterM;
+            }
+        }
+
+        /* Clamp the mel energies. */
+        constexpr float maxDb = 80.0;
+        const float clampLevelLowdB = maxMelEnergy - maxDb;
+        for (float& melEnergy : melEnergies) {
+            melEnergy = std::max(melEnergy, clampLevelLowdB);
+        }
+    }
+
+    std::vector<float> Wav2LetterMFCC::CreateDCTMatrix(
+                                        const int32_t inputLength,
+                                        const int32_t coefficientCount)
+    {
+        std::vector<float> dctMatix(inputLength * coefficientCount);
+
+        /* Orthonormal normalization. */
+        const float normalizerK0 = 2 * math::MathUtils::SqrtF32(1.0f /
+                                        static_cast<float>(4*inputLength));
+        const float normalizer = 2 * math::MathUtils::SqrtF32(1.0f /
+                                        static_cast<float>(2*inputLength));
+
+        const float angleIncr = M_PI / inputLength;
+        float angle = angleIncr;  /* We start using it at k = 1 loop. */
+
+        /* First row of DCT will use normalizer K0. */
+        for (int32_t n = 0; n < inputLength; ++n) {
+            dctMatix[n] = normalizerK0  /* cos(0) = 1 */;
+        }
+
+        /* Second row (index = 1) onwards, we use standard normalizer. */
+        for (int32_t k = 1, m = inputLength; k < coefficientCount; ++k, m += inputLength) {
+            for (int32_t n = 0; n < inputLength; ++n) {
+                dctMatix[m+n] = normalizer *
+                    math::MathUtils::CosineF32((n + 0.5f) * angle);
+            }
+            angle += angleIncr;
+        }
+        return dctMatix;
+    }
+
+    float Wav2LetterMFCC::GetMelFilterBankNormaliser(
+                                    const float&    leftMel,
+                                    const float&    rightMel,
+                                    const bool      useHTKMethod)
+    {
+        /* Slaney normalization for mel weights. */
+        return (2.0f / (MFCC::InverseMelScale(rightMel, useHTKMethod) -
+                MFCC::InverseMelScale(leftMel, useHTKMethod)));
+    }
+
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
diff --git a/source/application/api/use_case/asr/src/Wav2LetterModel.cc b/source/application/api/use_case/asr/src/Wav2LetterModel.cc
new file mode 100644
index 0000000..7b1e521
--- /dev/null
+++ b/source/application/api/use_case/asr/src/Wav2LetterModel.cc
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "Wav2LetterModel.hpp"
+
+#include "log_macros.h"
+
+
+const tflite::MicroOpResolver& arm::app::Wav2LetterModel::GetOpResolver()
+{
+    return this->m_opResolver;
+}
+
+bool arm::app::Wav2LetterModel::EnlistOperations()
+{
+    this->m_opResolver.AddConv2D();
+    this->m_opResolver.AddReshape();
+    this->m_opResolver.AddLeakyRelu();
+    this->m_opResolver.AddSoftmax();
+
+    if (kTfLiteOk == this->m_opResolver.AddEthosU()) {
+        info("Added %s support to op resolver\n",
+            tflite::GetString_ETHOSU());
+    } else {
+        printf_err("Failed to add Arm NPU support to op resolver.");
+        return false;
+    }
+    return true;
+}
diff --git a/source/application/api/use_case/asr/src/Wav2LetterPostprocess.cc b/source/application/api/use_case/asr/src/Wav2LetterPostprocess.cc
new file mode 100644
index 0000000..00e689b
--- /dev/null
+++ b/source/application/api/use_case/asr/src/Wav2LetterPostprocess.cc
@@ -0,0 +1,214 @@
+/*
+ * Copyright (c) 2021-2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "Wav2LetterPostprocess.hpp"
+
+#include "Wav2LetterModel.hpp"
+#include "log_macros.h"
+
+#include <cmath>
+
+namespace arm {
+namespace app {
+
+    AsrPostProcess::AsrPostProcess(TfLiteTensor* outputTensor, AsrClassifier& classifier,
+            const std::vector<std::string>& labels, std::vector<ClassificationResult>& results,
+            const uint32_t outputContextLen,
+            const uint32_t blankTokenIdx, const uint32_t reductionAxisIdx
+            ):
+            m_classifier(classifier),
+            m_outputTensor(outputTensor),
+            m_labels{labels},
+            m_results(results),
+            m_outputContextLen(outputContextLen),
+            m_countIterations(0),
+            m_blankTokenIdx(blankTokenIdx),
+            m_reductionAxisIdx(reductionAxisIdx)
+    {
+        this->m_outputInnerLen = AsrPostProcess::GetOutputInnerLen(this->m_outputTensor, this->m_outputContextLen);
+        this->m_totalLen = (2 * this->m_outputContextLen + this->m_outputInnerLen);
+    }
+
+    bool AsrPostProcess::DoPostProcess()
+    {
+        /* Basic checks. */
+        if (!this->IsInputValid(this->m_outputTensor, this->m_reductionAxisIdx)) {
+            return false;
+        }
+
+        /* Irrespective of tensor type, we use unsigned "byte" */
+        auto* ptrData = tflite::GetTensorData<uint8_t>(this->m_outputTensor);
+        const uint32_t elemSz = AsrPostProcess::GetTensorElementSize(this->m_outputTensor);
+
+        /* Other sanity checks. */
+        if (0 == elemSz) {
+            printf_err("Tensor type not supported for post processing\n");
+            return false;
+        } else if (elemSz * this->m_totalLen > this->m_outputTensor->bytes) {
+            printf_err("Insufficient number of tensor bytes\n");
+            return false;
+        }
+
+        /* Which axis do we need to process? */
+        switch (this->m_reductionAxisIdx) {
+            case Wav2LetterModel::ms_outputRowsIdx:
+                this->EraseSectionsRowWise(
+                        ptrData, elemSz * this->m_outputTensor->dims->data[Wav2LetterModel::ms_outputColsIdx],
+                        this->m_lastIteration);
+                break;
+            default:
+                printf_err("Unsupported axis index: %" PRIu32 "\n", this->m_reductionAxisIdx);
+                return false;
+        }
+        this->m_classifier.GetClassificationResults(this->m_outputTensor,
+                this->m_results, this->m_labels, 1);
+
+        return true;
+    }
+
+    bool AsrPostProcess::IsInputValid(TfLiteTensor* tensor, const uint32_t axisIdx) const
+    {
+        if (nullptr == tensor) {
+            return false;
+        }
+
+        if (static_cast<int>(axisIdx) >= tensor->dims->size) {
+            printf_err("Invalid axis index: %" PRIu32 "; Max: %d\n",
+                axisIdx, tensor->dims->size);
+            return false;
+        }
+
+        if (static_cast<int>(this->m_totalLen) !=
+                             tensor->dims->data[axisIdx]) {
+            printf_err("Unexpected tensor dimension for axis %" PRIu32", got %d.\n",
+                axisIdx, tensor->dims->data[axisIdx]);
+            return false;
+        }
+
+        return true;
+    }
+
+    uint32_t AsrPostProcess::GetTensorElementSize(TfLiteTensor* tensor)
+    {
+        switch(tensor->type) {
+            case kTfLiteUInt8:
+            case kTfLiteInt8:
+                return 1;
+            case kTfLiteInt16:
+                return 2;
+            case kTfLiteInt32:
+            case kTfLiteFloat32:
+                return 4;
+            default:
+                printf_err("Unsupported tensor type %s\n",
+                    TfLiteTypeGetName(tensor->type));
+        }
+
+        return 0;
+    }
+
+    bool AsrPostProcess::EraseSectionsRowWise(
+            uint8_t*         ptrData,
+            const uint32_t   strideSzBytes,
+            const bool       lastIteration)
+    {
+        /* In this case, the "zero-ing" is quite simple as the region
+         * to be zeroed sits in contiguous memory (row-major). */
+        const uint32_t eraseLen = strideSzBytes * this->m_outputContextLen;
+
+        /* Erase left context? */
+        if (this->m_countIterations > 0) {
+            /* Set output of each classification window to the blank token. */
+            std::memset(ptrData, 0, eraseLen);
+            for (size_t windowIdx = 0; windowIdx < this->m_outputContextLen; windowIdx++) {
+                ptrData[windowIdx*strideSzBytes + this->m_blankTokenIdx] = 1;
+            }
+        }
+
+        /* Erase right context? */
+        if (false == lastIteration) {
+            uint8_t* rightCtxPtr = ptrData + (strideSzBytes * (this->m_outputContextLen + this->m_outputInnerLen));
+            /* Set output of each classification window to the blank token. */
+            std::memset(rightCtxPtr, 0, eraseLen);
+            for (size_t windowIdx = 0; windowIdx < this->m_outputContextLen; windowIdx++) {
+                rightCtxPtr[windowIdx*strideSzBytes + this->m_blankTokenIdx] = 1;
+            }
+        }
+
+        if (lastIteration) {
+            this->m_countIterations = 0;
+        } else {
+            ++this->m_countIterations;
+        }
+
+        return true;
+    }
+
+    uint32_t AsrPostProcess::GetNumFeatureVectors(const Model& model)
+    {
+        TfLiteTensor* inputTensor = model.GetInputTensor(0);
+        const int inputRows = std::max(inputTensor->dims->data[Wav2LetterModel::ms_inputRowsIdx], 0);
+        if (inputRows == 0) {
+            printf_err("Error getting number of input rows for axis: %" PRIu32 "\n",
+                    Wav2LetterModel::ms_inputRowsIdx);
+        }
+        return inputRows;
+    }
+
+    uint32_t AsrPostProcess::GetOutputInnerLen(const TfLiteTensor* outputTensor, const uint32_t outputCtxLen)
+    {
+        const uint32_t outputRows = std::max(outputTensor->dims->data[Wav2LetterModel::ms_outputRowsIdx], 0);
+        if (outputRows == 0) {
+            printf_err("Error getting number of output rows for axis: %" PRIu32 "\n",
+                    Wav2LetterModel::ms_outputRowsIdx);
+        }
+
+        /* Watching for underflow. */
+        int innerLen = (outputRows - (2 * outputCtxLen));
+
+        return std::max(innerLen, 0);
+    }
+
+    uint32_t AsrPostProcess::GetOutputContextLen(const Model& model, const uint32_t inputCtxLen)
+    {
+        const uint32_t inputRows = AsrPostProcess::GetNumFeatureVectors(model);
+        const uint32_t inputInnerLen = inputRows - (2 * inputCtxLen);
+        constexpr uint32_t ms_outputRowsIdx = Wav2LetterModel::ms_outputRowsIdx;
+
+        /* Check to make sure that the input tensor supports the above
+         * context and inner lengths. */
+        if (inputRows <= 2 * inputCtxLen || inputRows <= inputInnerLen) {
+            printf_err("Input rows not compatible with ctx of %" PRIu32 "\n",
+                       inputCtxLen);
+            return 0;
+        }
+
+        TfLiteTensor* outputTensor = model.GetOutputTensor(0);
+        const uint32_t outputRows = std::max(outputTensor->dims->data[ms_outputRowsIdx], 0);
+        if (outputRows == 0) {
+            printf_err("Error getting number of output rows for axis: %" PRIu32 "\n",
+                       Wav2LetterModel::ms_outputRowsIdx);
+            return 0;
+        }
+
+        const float inOutRowRatio = static_cast<float>(inputRows) /
+                                     static_cast<float>(outputRows);
+
+        return std::round(static_cast<float>(inputCtxLen) / inOutRowRatio);
+    }
+
+} /* namespace app */
+} /* namespace arm */
diff --git a/source/application/api/use_case/asr/src/Wav2LetterPreprocess.cc b/source/application/api/use_case/asr/src/Wav2LetterPreprocess.cc
new file mode 100644
index 0000000..92b0631
--- /dev/null
+++ b/source/application/api/use_case/asr/src/Wav2LetterPreprocess.cc
@@ -0,0 +1,208 @@
+/*
+ * Copyright (c) 2021-2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "Wav2LetterPreprocess.hpp"
+
+#include "PlatformMath.hpp"
+#include "TensorFlowLiteMicro.hpp"
+
+#include <algorithm>
+#include <cmath>
+
+namespace arm {
+namespace app {
+
+    AsrPreProcess::AsrPreProcess(TfLiteTensor* inputTensor, const uint32_t numMfccFeatures,
+                                 const uint32_t numFeatureFrames, const uint32_t mfccWindowLen,
+                                 const uint32_t mfccWindowStride
+            ):
+            m_mfcc(numMfccFeatures, mfccWindowLen),
+            m_inputTensor(inputTensor),
+            m_mfccBuf(numMfccFeatures, numFeatureFrames),
+            m_delta1Buf(numMfccFeatures, numFeatureFrames),
+            m_delta2Buf(numMfccFeatures, numFeatureFrames),
+            m_mfccWindowLen(mfccWindowLen),
+            m_mfccWindowStride(mfccWindowStride),
+            m_numMfccFeats(numMfccFeatures),
+            m_numFeatureFrames(numFeatureFrames)
+    {
+        if (numMfccFeatures > 0 && mfccWindowLen > 0) {
+            this->m_mfcc.Init();
+        }
+    }
+
+    bool AsrPreProcess::DoPreProcess(const void* audioData, const size_t audioDataLen)
+    {
+        this->m_mfccSlidingWindow = audio::SlidingWindow<const int16_t>(
+                static_cast<const int16_t*>(audioData), audioDataLen,
+                this->m_mfccWindowLen, this->m_mfccWindowStride);
+
+        uint32_t mfccBufIdx = 0;
+
+        std::fill(m_mfccBuf.begin(), m_mfccBuf.end(), 0.f);
+        std::fill(m_delta1Buf.begin(), m_delta1Buf.end(), 0.f);
+        std::fill(m_delta2Buf.begin(), m_delta2Buf.end(), 0.f);
+
+        /* While we can slide over the audio. */
+        while (this->m_mfccSlidingWindow.HasNext()) {
+            const int16_t* mfccWindow = this->m_mfccSlidingWindow.Next();
+            auto mfccAudioData = std::vector<int16_t>(
+                                        mfccWindow,
+                                        mfccWindow + this->m_mfccWindowLen);
+            auto mfcc = this->m_mfcc.MfccCompute(mfccAudioData);
+            for (size_t i = 0; i < this->m_mfccBuf.size(0); ++i) {
+                this->m_mfccBuf(i, mfccBufIdx) = mfcc[i];
+            }
+            ++mfccBufIdx;
+        }
+
+        /* Pad MFCC if needed by adding MFCC for zeros. */
+        if (mfccBufIdx != this->m_numFeatureFrames) {
+            std::vector<int16_t> zerosWindow = std::vector<int16_t>(this->m_mfccWindowLen, 0);
+            std::vector<float> mfccZeros = this->m_mfcc.MfccCompute(zerosWindow);
+
+            while (mfccBufIdx != this->m_numFeatureFrames) {
+                memcpy(&this->m_mfccBuf(0, mfccBufIdx),
+                       mfccZeros.data(), sizeof(float) * m_numMfccFeats);
+                ++mfccBufIdx;
+            }
+        }
+
+        /* Compute first and second order deltas from MFCCs. */
+        AsrPreProcess::ComputeDeltas(this->m_mfccBuf, this->m_delta1Buf, this->m_delta2Buf);
+
+        /* Standardize calculated features. */
+        this->Standarize();
+
+        /* Quantise. */
+        QuantParams quantParams = GetTensorQuantParams(this->m_inputTensor);
+
+        if (0 == quantParams.scale) {
+            printf_err("Quantisation scale can't be 0\n");
+            return false;
+        }
+
+        switch(this->m_inputTensor->type) {
+            case kTfLiteUInt8:
+                return this->Quantise<uint8_t>(
+                        tflite::GetTensorData<uint8_t>(this->m_inputTensor), this->m_inputTensor->bytes,
+                        quantParams.scale, quantParams.offset);
+            case kTfLiteInt8:
+                return this->Quantise<int8_t>(
+                        tflite::GetTensorData<int8_t>(this->m_inputTensor), this->m_inputTensor->bytes,
+                        quantParams.scale, quantParams.offset);
+            default:
+                printf_err("Unsupported tensor type %s\n",
+                    TfLiteTypeGetName(this->m_inputTensor->type));
+        }
+
+        return false;
+    }
+
+    bool AsrPreProcess::ComputeDeltas(Array2d<float>& mfcc,
+                                      Array2d<float>& delta1,
+                                      Array2d<float>& delta2)
+    {
+        const std::vector <float> delta1Coeffs =
+            {6.66666667e-02,  5.00000000e-02,  3.33333333e-02,
+             1.66666667e-02, -3.46944695e-18, -1.66666667e-02,
+            -3.33333333e-02, -5.00000000e-02, -6.66666667e-02};
+
+        const std::vector <float> delta2Coeffs =
+            {0.06060606,      0.01515152,     -0.01731602,
+            -0.03679654,     -0.04329004,     -0.03679654,
+            -0.01731602,      0.01515152,      0.06060606};
+
+        if (delta1.size(0) == 0 || delta2.size(0) != delta1.size(0) ||
+            mfcc.size(0) == 0 || mfcc.size(1) == 0) {
+            return false;
+        }
+
+        /* Get the middle index; coeff vec len should always be odd. */
+        const size_t coeffLen = delta1Coeffs.size();
+        const size_t fMidIdx = (coeffLen - 1)/2;
+        const size_t numFeatures = mfcc.size(0);
+        const size_t numFeatVectors = mfcc.size(1);
+
+        /* Iterate through features in MFCC vector. */
+        for (size_t i = 0; i < numFeatures; ++i) {
+            /* For each feature, iterate through time (t) samples representing feature evolution and
+             * calculate d/dt and d^2/dt^2, using 1D convolution with differential kernels.
+             * Convolution padding = valid, result size is `time length - kernel length + 1`.
+             * The result is padded with 0 from both sides to match the size of initial time samples data.
+             *
+             * For the small filter, conv1D implementation as a simple loop is efficient enough.
+             * Filters of a greater size would need CMSIS-DSP functions to be used, like arm_fir_f32.
+             */
+
+            for (size_t j = fMidIdx; j < numFeatVectors - fMidIdx; ++j) {
+                float d1 = 0;
+                float d2 = 0;
+                const size_t mfccStIdx = j - fMidIdx;
+
+                for (size_t k = 0, m = coeffLen - 1; k < coeffLen; ++k, --m) {
+
+                    d1 +=  mfcc(i,mfccStIdx + k) * delta1Coeffs[m];
+                    d2 +=  mfcc(i,mfccStIdx + k) * delta2Coeffs[m];
+                }
+
+                delta1(i,j) = d1;
+                delta2(i,j) = d2;
+            }
+        }
+
+        return true;
+    }
+
+    void AsrPreProcess::StandardizeVecF32(Array2d<float>& vec)
+    {
+        auto mean = math::MathUtils::MeanF32(vec.begin(), vec.totalSize());
+        auto stddev = math::MathUtils::StdDevF32(vec.begin(), vec.totalSize(), mean);
+
+        debug("Mean: %f, Stddev: %f\n", mean, stddev);
+        if (stddev == 0) {
+            std::fill(vec.begin(), vec.end(), 0);
+        } else {
+            const float stddevInv = 1.f/stddev;
+            const float normalisedMean = mean/stddev;
+
+            auto NormalisingFunction = [=](float& value) {
+                value = value * stddevInv - normalisedMean;
+            };
+            std::for_each(vec.begin(), vec.end(), NormalisingFunction);
+        }
+    }
+
+    void AsrPreProcess::Standarize()
+    {
+        AsrPreProcess::StandardizeVecF32(this->m_mfccBuf);
+        AsrPreProcess::StandardizeVecF32(this->m_delta1Buf);
+        AsrPreProcess::StandardizeVecF32(this->m_delta2Buf);
+    }
+
+    float AsrPreProcess::GetQuantElem(
+                const float     elem,
+                const float     quantScale,
+                const int       quantOffset,
+                const float     minVal,
+                const float     maxVal)
+    {
+        float val = std::round((elem/quantScale) + quantOffset);
+        return std::min<float>(std::max<float>(val, minVal), maxVal);
+    }
+
+} /* namespace app */
+} /* namespace arm */
\ No newline at end of file
diff --git a/source/application/api/use_case/img_class/CMakeLists.txt b/source/application/api/use_case/img_class/CMakeLists.txt
new file mode 100644
index 0000000..f4818d8
--- /dev/null
+++ b/source/application/api/use_case/img_class/CMakeLists.txt
@@ -0,0 +1,39 @@
+#----------------------------------------------------------------------------
+#  Copyright (c) 2022 Arm Limited. All rights reserved.
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#----------------------------------------------------------------------------
+#########################################################
+#                IMG CLASS API library                  #
+#########################################################
+cmake_minimum_required(VERSION 3.15.6)
+
+set(IMG_CLASS_API_TARGET img_class_api)
+project(${IMG_CLASS_API_TARGET}
+        DESCRIPTION     "Image classification use case API library"
+        LANGUAGES       C CXX)
+
+# Create static library
+add_library(${IMG_CLASS_API_TARGET} STATIC
+    src/ImgClassProcessing.cc
+    src/MobileNetModel.cc)
+
+target_include_directories(${IMG_CLASS_API_TARGET} PUBLIC include)
+
+target_link_libraries(${IMG_CLASS_API_TARGET} PUBLIC common_api)
+
+message(STATUS "*******************************************************")
+message(STATUS "Library                                : " ${IMG_CLASS_API_TARGET})
+message(STATUS "CMAKE_SYSTEM_PROCESSOR                 : " ${CMAKE_SYSTEM_PROCESSOR})
+message(STATUS "*******************************************************")
diff --git a/source/application/api/use_case/img_class/include/ImgClassProcessing.hpp b/source/application/api/use_case/img_class/include/ImgClassProcessing.hpp
new file mode 100644
index 0000000..55b5ce1
--- /dev/null
+++ b/source/application/api/use_case/img_class/include/ImgClassProcessing.hpp
@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef IMG_CLASS_PROCESSING_HPP
+#define IMG_CLASS_PROCESSING_HPP
+
+#include "BaseProcessing.hpp"
+#include "Classifier.hpp"
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   Pre-processing class for Image Classification use case.
+     *          Implements methods declared by BasePreProcess and anything else needed
+     *          to populate input tensors ready for inference.
+     */
+    class ImgClassPreProcess : public BasePreProcess {
+
+    public:
+        /**
+         * @brief       Constructor
+         * @param[in]   inputTensor     Pointer to the TFLite Micro input Tensor.
+         * @param[in]   convertToInt8   Should the image be converted to Int8 range.
+         **/
+        explicit ImgClassPreProcess(TfLiteTensor* inputTensor, bool convertToInt8);
+
+        /**
+         * @brief       Should perform pre-processing of 'raw' input image data and load it into
+         *              TFLite Micro input tensors ready for inference
+         * @param[in]   input      Pointer to the data that pre-processing will work on.
+         * @param[in]   inputSize  Size of the input data.
+         * @return      true if successful, false otherwise.
+         **/
+        bool DoPreProcess(const void* input, size_t inputSize) override;
+
+    private:
+        TfLiteTensor* m_inputTensor;
+        bool m_convertToInt8;
+    };
+
+    /**
+     * @brief   Post-processing class for Image Classification use case.
+     *          Implements methods declared by BasePostProcess and anything else needed
+     *          to populate result vector.
+     */
+    class ImgClassPostProcess : public BasePostProcess {
+
+    public:
+        /**
+         * @brief       Constructor
+         * @param[in]   outputTensor  Pointer to the TFLite Micro output Tensor.
+         * @param[in]   classifier    Classifier object used to get top N results from classification.
+         * @param[in]   labels        Vector of string labels to identify each output of the model.
+         * @param[in]   results       Vector of classification results to store decoded outputs.
+         **/
+        ImgClassPostProcess(TfLiteTensor* outputTensor, Classifier& classifier,
+                            const std::vector<std::string>& labels,
+                            std::vector<ClassificationResult>& results);
+
+        /**
+         * @brief       Should perform post-processing of the result of inference then
+         *              populate classification result data for any later use.
+         * @return      true if successful, false otherwise.
+         **/
+        bool DoPostProcess() override;
+
+    private:
+        TfLiteTensor* m_outputTensor;
+        Classifier& m_imgClassifier;
+        const std::vector<std::string>& m_labels;
+        std::vector<ClassificationResult>& m_results;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* IMG_CLASS_PROCESSING_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/img_class/include/MobileNetModel.hpp b/source/application/api/use_case/img_class/include/MobileNetModel.hpp
new file mode 100644
index 0000000..adaa9c2
--- /dev/null
+++ b/source/application/api/use_case/img_class/include/MobileNetModel.hpp
@@ -0,0 +1,51 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef IMG_CLASS_MOBILENETMODEL_HPP
+#define IMG_CLASS_MOBILENETMODEL_HPP
+
+#include "Model.hpp"
+
+namespace arm {
+namespace app {
+
+    class MobileNetModel : public Model {
+
+    public:
+        /* Indices for the expected model - based on input tensor shape */
+        static constexpr uint32_t ms_inputRowsIdx     = 1;
+        static constexpr uint32_t ms_inputColsIdx     = 2;
+        static constexpr uint32_t ms_inputChannelsIdx = 3;
+
+    protected:
+        /** @brief   Gets the reference to op resolver interface class. */
+        const tflite::MicroOpResolver& GetOpResolver() override;
+
+        /** @brief   Adds operations to the op resolver instance. */
+        bool EnlistOperations() override;
+
+    private:
+        /* Maximum number of individual operations that can be enlisted. */
+        static constexpr int ms_maxOpCnt = 7;
+
+        /* A mutable op resolver instance. */
+        tflite::MicroMutableOpResolver<ms_maxOpCnt> m_opResolver;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* IMG_CLASS_MOBILENETMODEL_HPP */
diff --git a/source/application/api/use_case/img_class/src/ImgClassProcessing.cc b/source/application/api/use_case/img_class/src/ImgClassProcessing.cc
new file mode 100644
index 0000000..491e751
--- /dev/null
+++ b/source/application/api/use_case/img_class/src/ImgClassProcessing.cc
@@ -0,0 +1,66 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "ImgClassProcessing.hpp"
+
+#include "ImageUtils.hpp"
+#include "log_macros.h"
+
+namespace arm {
+namespace app {
+
+    ImgClassPreProcess::ImgClassPreProcess(TfLiteTensor* inputTensor, bool convertToInt8)
+    :m_inputTensor{inputTensor},
+     m_convertToInt8{convertToInt8}
+    {}
+
+    bool ImgClassPreProcess::DoPreProcess(const void* data, size_t inputSize)
+    {
+        if (data == nullptr) {
+            printf_err("Data pointer is null");
+            return false;
+        }
+
+        auto input = static_cast<const uint8_t*>(data);
+
+        std::memcpy(this->m_inputTensor->data.data, input, inputSize);
+        debug("Input tensor populated \n");
+
+        if (this->m_convertToInt8) {
+            image::ConvertImgToInt8(this->m_inputTensor->data.data, this->m_inputTensor->bytes);
+        }
+
+        return true;
+    }
+
+    ImgClassPostProcess::ImgClassPostProcess(TfLiteTensor* outputTensor, Classifier& classifier,
+                                             const std::vector<std::string>& labels,
+                                             std::vector<ClassificationResult>& results)
+            :m_outputTensor{outputTensor},
+             m_imgClassifier{classifier},
+             m_labels{labels},
+             m_results{results}
+    {}
+
+    bool ImgClassPostProcess::DoPostProcess()
+    {
+        return this->m_imgClassifier.GetClassificationResults(
+                this->m_outputTensor, this->m_results,
+                this->m_labels, 5, false);
+    }
+
+} /* namespace app */
+} /* namespace arm */
\ No newline at end of file
diff --git a/source/application/api/use_case/img_class/src/MobileNetModel.cc b/source/application/api/use_case/img_class/src/MobileNetModel.cc
new file mode 100644
index 0000000..b700d70
--- /dev/null
+++ b/source/application/api/use_case/img_class/src/MobileNetModel.cc
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "MobileNetModel.hpp"
+#include "log_macros.h"
+
+const tflite::MicroOpResolver& arm::app::MobileNetModel::GetOpResolver()
+{
+    return this->m_opResolver;
+}
+
+bool arm::app::MobileNetModel::EnlistOperations()
+{
+    this->m_opResolver.AddDepthwiseConv2D();
+    this->m_opResolver.AddConv2D();
+    this->m_opResolver.AddAveragePool2D();
+    this->m_opResolver.AddAdd();
+    this->m_opResolver.AddReshape();
+    this->m_opResolver.AddSoftmax();
+
+    if (kTfLiteOk == this->m_opResolver.AddEthosU()) {
+        info("Added %s support to op resolver\n",
+            tflite::GetString_ETHOSU());
+    } else {
+        printf_err("Failed to add Arm NPU support to op resolver.");
+        return false;
+    }
+    return true;
+}
diff --git a/source/application/api/use_case/inference_runner/CMakeLists.txt b/source/application/api/use_case/inference_runner/CMakeLists.txt
new file mode 100644
index 0000000..d0fe629
--- /dev/null
+++ b/source/application/api/use_case/inference_runner/CMakeLists.txt
@@ -0,0 +1,37 @@
+#----------------------------------------------------------------------------
+#  Copyright (c) 2022 Arm Limited. All rights reserved.
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#----------------------------------------------------------------------------
+#########################################################
+#              INFERENCE RUNNER API library             #
+#########################################################
+cmake_minimum_required(VERSION 3.15.6)
+
+set(INFERENCE_RUNNER_API_TARGET inference_runner_api)
+project(${INFERENCE_RUNNER_API_TARGET}
+        DESCRIPTION     "Inference runner use case API library"
+        LANGUAGES       C CXX)
+
+# Create static library
+add_library(${INFERENCE_RUNNER_API_TARGET} STATIC src/TestModel.cc)
+
+target_include_directories(${INFERENCE_RUNNER_API_TARGET} PUBLIC include)
+
+target_link_libraries(${INFERENCE_RUNNER_API_TARGET} PUBLIC common_api)
+
+message(STATUS "*******************************************************")
+message(STATUS "Library                                : " ${INFERENCE_RUNNER_API_TARGET})
+message(STATUS "CMAKE_SYSTEM_PROCESSOR                 : " ${CMAKE_SYSTEM_PROCESSOR})
+message(STATUS "*******************************************************")
diff --git a/source/application/api/use_case/inference_runner/include/TestModel.hpp b/source/application/api/use_case/inference_runner/include/TestModel.hpp
new file mode 100644
index 0000000..648198c
--- /dev/null
+++ b/source/application/api/use_case/inference_runner/include/TestModel.hpp
@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef INF_RUNNER_TESTMODEL_HPP
+#define INF_RUNNER_TESTMODEL_HPP
+
+#include "Model.hpp"
+
+namespace arm {
+namespace app {
+
+    class TestModel : public Model {
+
+    protected:
+        /** @brief   Gets the reference to op resolver interface class. */
+        const tflite::AllOpsResolver& GetOpResolver() override;
+
+        /** @brief   Adds operations to the op resolver instance, not needed as using AllOpsResolver. */
+        bool EnlistOperations() override {return false;}
+
+    private:
+
+        /* No need to define individual ops at the cost of extra memory. */
+        tflite::AllOpsResolver m_opResolver;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* INF_RUNNER_TESTMODEL_HPP */
diff --git a/source/application/api/use_case/inference_runner/src/TestModel.cc b/source/application/api/use_case/inference_runner/src/TestModel.cc
new file mode 100644
index 0000000..1891e44
--- /dev/null
+++ b/source/application/api/use_case/inference_runner/src/TestModel.cc
@@ -0,0 +1,23 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "TestModel.hpp"
+#include "log_macros.h"
+
+const tflite::AllOpsResolver& arm::app::TestModel::GetOpResolver()
+{
+    return this->m_opResolver;
+}
diff --git a/source/application/api/use_case/kws/CMakeLists.txt b/source/application/api/use_case/kws/CMakeLists.txt
new file mode 100644
index 0000000..3256d03
--- /dev/null
+++ b/source/application/api/use_case/kws/CMakeLists.txt
@@ -0,0 +1,39 @@
+#----------------------------------------------------------------------------
+#  Copyright (c) 2022 Arm Limited. All rights reserved.
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#----------------------------------------------------------------------------
+#########################################################
+#            KEYWORD SPOTTING API library               #
+#########################################################
+cmake_minimum_required(VERSION 3.15.6)
+
+set(KWS_API_TARGET kws_api)
+project(${KWS_API_TARGET}
+        DESCRIPTION     "Keyword spotting use case API library"
+        LANGUAGES       C CXX)
+
+# Create static library
+add_library(${KWS_API_TARGET} STATIC
+    src/KwsProcessing.cc
+    src/MicroNetKwsModel.cc)
+
+target_include_directories(${KWS_API_TARGET} PUBLIC include)
+
+target_link_libraries(${KWS_API_TARGET} PUBLIC common_api)
+
+message(STATUS "*******************************************************")
+message(STATUS "Library                                : " ${KWS_API_TARGET})
+message(STATUS "CMAKE_SYSTEM_PROCESSOR                 : " ${CMAKE_SYSTEM_PROCESSOR})
+message(STATUS "*******************************************************")
diff --git a/source/application/api/use_case/kws/include/KwsProcessing.hpp b/source/application/api/use_case/kws/include/KwsProcessing.hpp
new file mode 100644
index 0000000..0ede425
--- /dev/null
+++ b/source/application/api/use_case/kws/include/KwsProcessing.hpp
@@ -0,0 +1,137 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef KWS_PROCESSING_HPP
+#define KWS_PROCESSING_HPP
+
+#include "AudioUtils.hpp"
+#include "BaseProcessing.hpp"
+#include "Classifier.hpp"
+#include "MicroNetKwsMfcc.hpp"
+
+#include <functional>
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   Pre-processing class for Keyword Spotting use case.
+     *          Implements methods declared by BasePreProcess and anything else needed
+     *          to populate input tensors ready for inference.
+     */
+    class KwsPreProcess : public BasePreProcess {
+
+    public:
+        /**
+         * @brief       Constructor
+         * @param[in]   inputTensor        Pointer to the TFLite Micro input Tensor.
+         * @param[in]   numFeatures        How many MFCC features to use.
+         * @param[in]   numFeatureFrames   Number of MFCC vectors that need to be calculated
+         *                                 for an inference.
+         * @param[in]   mfccFrameLength    Number of audio samples used to calculate one set of MFCC values when
+         *                                 sliding a window through the audio sample.
+         * @param[in]   mfccFrameStride    Number of audio samples between consecutive windows.
+         **/
+        explicit KwsPreProcess(TfLiteTensor* inputTensor, size_t numFeatures, size_t numFeatureFrames,
+                               int mfccFrameLength, int mfccFrameStride);
+
+        /**
+         * @brief       Should perform pre-processing of 'raw' input audio data and load it into
+         *              TFLite Micro input tensors ready for inference.
+         * @param[in]   input      Pointer to the data that pre-processing will work on.
+         * @param[in]   inputSize  Size of the input data.
+         * @return      true if successful, false otherwise.
+         **/
+        bool DoPreProcess(const void* input, size_t inputSize) override;
+
+        size_t m_audioWindowIndex = 0;  /* Index of audio slider, used when caching features in longer clips. */
+        size_t m_audioDataWindowSize;   /* Amount of audio needed for 1 inference. */
+        size_t m_audioDataStride;       /* Amount of audio to stride across if doing >1 inference in longer clips. */
+
+    private:
+        TfLiteTensor* m_inputTensor;    /* Model input tensor. */
+        const int m_mfccFrameLength;
+        const int m_mfccFrameStride;
+        const size_t m_numMfccFrames;   /* How many sets of m_numMfccFeats. */
+
+        audio::MicroNetKwsMFCC m_mfcc;
+        audio::SlidingWindow<const int16_t> m_mfccSlidingWindow;
+        size_t m_numMfccVectorsInAudioStride;
+        size_t m_numReusedMfccVectors;
+        std::function<void (std::vector<int16_t>&, int, bool, size_t)> m_mfccFeatureCalculator;
+
+        /**
+         * @brief Returns a function to perform feature calculation and populates input tensor data with
+         * MFCC data.
+         *
+         * Input tensor data type check is performed to choose correct MFCC feature data type.
+         * If tensor has an integer data type then original features are quantised.
+         *
+         * Warning: MFCC calculator provided as input must have the same life scope as returned function.
+         *
+         * @param[in]       mfcc          MFCC feature calculator.
+         * @param[in,out]   inputTensor   Input tensor pointer to store calculated features.
+         * @param[in]       cacheSize     Size of the feature vectors cache (number of feature vectors).
+         * @return          Function to be called providing audio sample and sliding window index.
+         */
+        std::function<void (std::vector<int16_t>&, int, bool, size_t)>
+        GetFeatureCalculator(audio::MicroNetKwsMFCC&  mfcc,
+                             TfLiteTensor*            inputTensor,
+                             size_t                   cacheSize);
+
+        template<class T>
+        std::function<void (std::vector<int16_t>&, size_t, bool, size_t)>
+        FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize,
+                    std::function<std::vector<T> (std::vector<int16_t>& )> compute);
+    };
+
+    /**
+     * @brief   Post-processing class for Keyword Spotting use case.
+     *          Implements methods declared by BasePostProcess and anything else needed
+     *          to populate result vector.
+     */
+    class KwsPostProcess : public BasePostProcess {
+
+    private:
+        TfLiteTensor* m_outputTensor;                   /* Model output tensor. */
+        Classifier& m_kwsClassifier;                    /* KWS Classifier object. */
+        const std::vector<std::string>& m_labels;       /* KWS Labels. */
+        std::vector<ClassificationResult>& m_results;   /* Results vector for a single inference. */
+
+    public:
+        /**
+         * @brief           Constructor
+         * @param[in]       outputTensor   Pointer to the TFLite Micro output Tensor.
+         * @param[in]       classifier     Classifier object used to get top N results from classification.
+         * @param[in]       labels         Vector of string labels to identify each output of the model.
+         * @param[in/out]   results        Vector of classification results to store decoded outputs.
+         **/
+        KwsPostProcess(TfLiteTensor* outputTensor, Classifier& classifier,
+                       const std::vector<std::string>& labels,
+                       std::vector<ClassificationResult>& results);
+
+        /**
+         * @brief    Should perform post-processing of the result of inference then
+         *           populate KWS result data for any later use.
+         * @return   true if successful, false otherwise.
+         **/
+        bool DoPostProcess() override;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* KWS_PROCESSING_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/kws/include/KwsResult.hpp b/source/application/api/use_case/kws/include/KwsResult.hpp
new file mode 100644
index 0000000..38f32b4
--- /dev/null
+++ b/source/application/api/use_case/kws/include/KwsResult.hpp
@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef KWS_RESULT_HPP
+#define KWS_RESULT_HPP
+
+#include "ClassificationResult.hpp"
+
+#include <vector>
+
+namespace arm {
+namespace app {
+namespace kws {
+
+    using ResultVec = std::vector<arm::app::ClassificationResult>;
+
+    /* Structure for holding kws result. */
+    class KwsResult {
+
+    public:
+        ResultVec       m_resultVec;        /* Container for "thresholded" classification results. */
+        float           m_timeStamp;        /* Audio timestamp for this result. */
+        uint32_t        m_inferenceNumber;  /* Corresponding inference number. */
+        float           m_threshold;        /* Threshold value for `m_resultVec`. */
+
+        KwsResult() = delete;
+        KwsResult(ResultVec&        resultVec,
+                  const float       timestamp,
+                  const uint32_t    inferenceIdx,
+                  const float       scoreThreshold) {
+
+            this->m_threshold = scoreThreshold;
+            this->m_timeStamp = timestamp;
+            this->m_inferenceNumber = inferenceIdx;
+
+            this->m_resultVec = ResultVec();
+            for (auto & i : resultVec) {
+                if (i.m_normalisedVal >= this->m_threshold) {
+                    this->m_resultVec.emplace_back(i);
+                }
+            }
+        }
+        ~KwsResult() = default;
+    };
+
+} /* namespace kws */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* KWS_RESULT_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/kws/include/MicroNetKwsMfcc.hpp b/source/application/api/use_case/kws/include/MicroNetKwsMfcc.hpp
new file mode 100644
index 0000000..b2565a3
--- /dev/null
+++ b/source/application/api/use_case/kws/include/MicroNetKwsMfcc.hpp
@@ -0,0 +1,50 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef KWS_MICRONET_MFCC_HPP
+#define KWS_MICRONET_MFCC_HPP
+
+#include "Mfcc.hpp"
+
+namespace arm {
+namespace app {
+namespace audio {
+
+    /* Class to provide MicroNet specific MFCC calculation requirements. */
+    class MicroNetKwsMFCC : public MFCC {
+
+    public:
+        static constexpr uint32_t  ms_defaultSamplingFreq = 16000;
+        static constexpr uint32_t  ms_defaultNumFbankBins =    40;
+        static constexpr uint32_t  ms_defaultMelLoFreq    =    20;
+        static constexpr uint32_t  ms_defaultMelHiFreq    =  4000;
+        static constexpr bool      ms_defaultUseHtkMethod =  true;
+
+        explicit MicroNetKwsMFCC(const size_t numFeats, const size_t frameLen)
+            :  MFCC(MfccParams(
+                        ms_defaultSamplingFreq, ms_defaultNumFbankBins,
+                        ms_defaultMelLoFreq, ms_defaultMelHiFreq,
+                        numFeats, frameLen, ms_defaultUseHtkMethod))
+        {}
+        MicroNetKwsMFCC()  = delete;
+        ~MicroNetKwsMFCC() = default;
+    };
+
+} /* namespace audio */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* KWS_MICRONET_MFCC_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/kws/include/MicroNetKwsModel.hpp b/source/application/api/use_case/kws/include/MicroNetKwsModel.hpp
new file mode 100644
index 0000000..3d2f3de
--- /dev/null
+++ b/source/application/api/use_case/kws/include/MicroNetKwsModel.hpp
@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef KWS_MICRONETMODEL_HPP
+#define KWS_MICRONETMODEL_HPP
+
+#include "Model.hpp"
+
+namespace arm {
+namespace app {
+namespace kws {
+    extern const int g_FrameLength;
+    extern const int g_FrameStride;
+    extern const float g_ScoreThreshold;
+    extern const uint32_t g_NumMfcc;
+    extern const uint32_t g_NumAudioWins;
+} /* namespace kws */
+} /* namespace app */
+} /* namespace arm */
+
+namespace arm {
+namespace app {
+
+    class MicroNetKwsModel : public Model {
+    public:
+        /* Indices for the expected model - based on input and output tensor shapes */
+        static constexpr uint32_t ms_inputRowsIdx = 1;
+        static constexpr uint32_t ms_inputColsIdx = 2;
+        static constexpr uint32_t ms_outputRowsIdx = 2;
+        static constexpr uint32_t ms_outputColsIdx = 3;
+
+    protected:
+        /** @brief   Gets the reference to op resolver interface class. */
+        const tflite::MicroOpResolver& GetOpResolver() override;
+
+        /** @brief   Adds operations to the op resolver instance. */
+        bool EnlistOperations() override;
+
+    private:
+        /* Maximum number of individual operations that can be enlisted. */
+        static constexpr int ms_maxOpCnt = 7;
+
+        /* A mutable op resolver instance. */
+        tflite::MicroMutableOpResolver<ms_maxOpCnt> m_opResolver;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* KWS_MICRONETMODEL_HPP */
diff --git a/source/application/api/use_case/kws/src/KwsProcessing.cc b/source/application/api/use_case/kws/src/KwsProcessing.cc
new file mode 100644
index 0000000..40de498
--- /dev/null
+++ b/source/application/api/use_case/kws/src/KwsProcessing.cc
@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "KwsProcessing.hpp"
+#include "log_macros.h"
+#include "MicroNetKwsModel.hpp"
+
+namespace arm {
+namespace app {
+
+    KwsPreProcess::KwsPreProcess(TfLiteTensor* inputTensor, size_t numFeatures, size_t numMfccFrames,
+            int mfccFrameLength, int mfccFrameStride
+        ):
+        m_inputTensor{inputTensor},
+        m_mfccFrameLength{mfccFrameLength},
+        m_mfccFrameStride{mfccFrameStride},
+        m_numMfccFrames{numMfccFrames},
+        m_mfcc{audio::MicroNetKwsMFCC(numFeatures, mfccFrameLength)}
+    {
+        this->m_mfcc.Init();
+
+        /* Deduce the data length required for 1 inference from the network parameters. */
+        this->m_audioDataWindowSize = this->m_numMfccFrames * this->m_mfccFrameStride +
+                (this->m_mfccFrameLength - this->m_mfccFrameStride);
+
+        /* Creating an MFCC feature sliding window for the data required for 1 inference. */
+        this->m_mfccSlidingWindow = audio::SlidingWindow<const int16_t>(nullptr, this->m_audioDataWindowSize,
+                this->m_mfccFrameLength, this->m_mfccFrameStride);
+
+        /* For longer audio clips we choose to move by half the audio window size
+         * => for a 1 second window size there is an overlap of 0.5 seconds. */
+        this->m_audioDataStride = this->m_audioDataWindowSize / 2;
+
+        /* To have the previously calculated features re-usable, stride must be multiple
+         * of MFCC features window stride. Reduce stride through audio if needed. */
+        if (0 != this->m_audioDataStride % this->m_mfccFrameStride) {
+            this->m_audioDataStride -= this->m_audioDataStride % this->m_mfccFrameStride;
+        }
+
+        this->m_numMfccVectorsInAudioStride = this->m_audioDataStride / this->m_mfccFrameStride;
+
+        /* Calculate number of the feature vectors in the window overlap region.
+         * These feature vectors will be reused.*/
+        this->m_numReusedMfccVectors = this->m_mfccSlidingWindow.TotalStrides() + 1
+                - this->m_numMfccVectorsInAudioStride;
+
+        /* Construct feature calculation function. */
+        this->m_mfccFeatureCalculator = GetFeatureCalculator(this->m_mfcc, this->m_inputTensor,
+                                                             this->m_numReusedMfccVectors);
+
+        if (!this->m_mfccFeatureCalculator) {
+            printf_err("Feature calculator not initialized.");
+        }
+    }
+
+    bool KwsPreProcess::DoPreProcess(const void* data, size_t inputSize)
+    {
+        UNUSED(inputSize);
+        if (data == nullptr) {
+            printf_err("Data pointer is null");
+        }
+
+        /* Set the features sliding window to the new address. */
+        auto input = static_cast<const int16_t*>(data);
+        this->m_mfccSlidingWindow.Reset(input);
+
+        /* Cache is only usable if we have more than 1 inference in an audio clip. */
+        bool useCache = this->m_audioWindowIndex > 0 && this->m_numReusedMfccVectors > 0;
+
+        /* Use a sliding window to calculate MFCC features frame by frame. */
+        while (this->m_mfccSlidingWindow.HasNext()) {
+            const int16_t* mfccWindow = this->m_mfccSlidingWindow.Next();
+
+            std::vector<int16_t> mfccFrameAudioData = std::vector<int16_t>(mfccWindow,
+                    mfccWindow + this->m_mfccFrameLength);
+
+            /* Compute features for this window and write them to input tensor. */
+            this->m_mfccFeatureCalculator(mfccFrameAudioData, this->m_mfccSlidingWindow.Index(),
+                                          useCache, this->m_numMfccVectorsInAudioStride);
+        }
+
+        debug("Input tensor populated \n");
+
+        return true;
+    }
+
+    /**
+     * @brief Generic feature calculator factory.
+     *
+     * Returns lambda function to compute features using features cache.
+     * Real features math is done by a lambda function provided as a parameter.
+     * Features are written to input tensor memory.
+     *
+     * @tparam T                Feature vector type.
+     * @param[in] inputTensor   Model input tensor pointer.
+     * @param[in] cacheSize     Number of feature vectors to cache. Defined by the sliding window overlap.
+     * @param[in] compute       Features calculator function.
+     * @return                  Lambda function to compute features.
+     */
+    template<class T>
+    std::function<void (std::vector<int16_t>&, size_t, bool, size_t)>
+    KwsPreProcess::FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize,
+                               std::function<std::vector<T> (std::vector<int16_t>& )> compute)
+    {
+        /* Feature cache to be captured by lambda function. */
+        static std::vector<std::vector<T>> featureCache = std::vector<std::vector<T>>(cacheSize);
+
+        return [=](std::vector<int16_t>& audioDataWindow,
+                   size_t index,
+                   bool useCache,
+                   size_t featuresOverlapIndex)
+        {
+            T* tensorData = tflite::GetTensorData<T>(inputTensor);
+            std::vector<T> features;
+
+            /* Reuse features from cache if cache is ready and sliding windows overlap.
+             * Overlap is in the beginning of sliding window with a size of a feature cache. */
+            if (useCache && index < featureCache.size()) {
+                features = std::move(featureCache[index]);
+            } else {
+                features = std::move(compute(audioDataWindow));
+            }
+            auto size = features.size();
+            auto sizeBytes = sizeof(T) * size;
+            std::memcpy(tensorData + (index * size), features.data(), sizeBytes);
+
+            /* Start renewing cache as soon iteration goes out of the windows overlap. */
+            if (index >= featuresOverlapIndex) {
+                featureCache[index - featuresOverlapIndex] = std::move(features);
+            }
+        };
+    }
+
+    template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)>
+    KwsPreProcess::FeatureCalc<int8_t>(TfLiteTensor* inputTensor,
+                                       size_t cacheSize,
+                                       std::function<std::vector<int8_t> (std::vector<int16_t>&)> compute);
+
+    template std::function<void(std::vector<int16_t>&, size_t, bool, size_t)>
+    KwsPreProcess::FeatureCalc<float>(TfLiteTensor* inputTensor,
+                                      size_t cacheSize,
+                                      std::function<std::vector<float>(std::vector<int16_t>&)> compute);
+
+
+    std::function<void (std::vector<int16_t>&, int, bool, size_t)>
+    KwsPreProcess::GetFeatureCalculator(audio::MicroNetKwsMFCC& mfcc, TfLiteTensor* inputTensor, size_t cacheSize)
+    {
+        std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> mfccFeatureCalc;
+
+        TfLiteQuantization quant = inputTensor->quantization;
+
+        if (kTfLiteAffineQuantization == quant.type) {
+            auto *quantParams = (TfLiteAffineQuantization *) quant.params;
+            const float quantScale = quantParams->scale->data[0];
+            const int quantOffset = quantParams->zero_point->data[0];
+
+            switch (inputTensor->type) {
+                case kTfLiteInt8: {
+                    mfccFeatureCalc = this->FeatureCalc<int8_t>(inputTensor,
+                                                          cacheSize,
+                                                          [=, &mfcc](std::vector<int16_t>& audioDataWindow) {
+                                                              return mfcc.MfccComputeQuant<int8_t>(audioDataWindow,
+                                                                                                   quantScale,
+                                                                                                   quantOffset);
+                                                          }
+                    );
+                    break;
+                }
+                default:
+                printf_err("Tensor type %s not supported\n", TfLiteTypeGetName(inputTensor->type));
+            }
+        } else {
+            mfccFeatureCalc = this->FeatureCalc<float>(inputTensor, cacheSize,
+                    [&mfcc](std::vector<int16_t>& audioDataWindow) {
+                return mfcc.MfccCompute(audioDataWindow); }
+                );
+        }
+        return mfccFeatureCalc;
+    }
+
+    KwsPostProcess::KwsPostProcess(TfLiteTensor* outputTensor, Classifier& classifier,
+                                   const std::vector<std::string>& labels,
+                                   std::vector<ClassificationResult>& results)
+            :m_outputTensor{outputTensor},
+             m_kwsClassifier{classifier},
+             m_labels{labels},
+             m_results{results}
+    {}
+
+    bool KwsPostProcess::DoPostProcess()
+    {
+        return this->m_kwsClassifier.GetClassificationResults(
+                this->m_outputTensor, this->m_results,
+                this->m_labels, 1, true);
+    }
+
+} /* namespace app */
+} /* namespace arm */
\ No newline at end of file
diff --git a/source/application/api/use_case/kws/src/MicroNetKwsModel.cc b/source/application/api/use_case/kws/src/MicroNetKwsModel.cc
new file mode 100644
index 0000000..bedca99
--- /dev/null
+++ b/source/application/api/use_case/kws/src/MicroNetKwsModel.cc
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "MicroNetKwsModel.hpp"
+#include "log_macros.h"
+
+const tflite::MicroOpResolver& arm::app::MicroNetKwsModel::GetOpResolver()
+{
+    return this->m_opResolver;
+}
+
+bool arm::app::MicroNetKwsModel::EnlistOperations()
+{
+    this->m_opResolver.AddReshape();
+    this->m_opResolver.AddAveragePool2D();
+    this->m_opResolver.AddConv2D();
+    this->m_opResolver.AddDepthwiseConv2D();
+    this->m_opResolver.AddFullyConnected();
+    this->m_opResolver.AddRelu();
+
+    if (kTfLiteOk == this->m_opResolver.AddEthosU()) {
+        info("Added %s support to op resolver\n",
+            tflite::GetString_ETHOSU());
+    } else {
+        printf_err("Failed to add Arm NPU support to op resolver.");
+        return false;
+    }
+    return true;
+}
diff --git a/source/application/api/use_case/noise_reduction/CMakeLists.txt b/source/application/api/use_case/noise_reduction/CMakeLists.txt
new file mode 100644
index 0000000..5fa9a73
--- /dev/null
+++ b/source/application/api/use_case/noise_reduction/CMakeLists.txt
@@ -0,0 +1,40 @@
+#----------------------------------------------------------------------------
+#  Copyright (c) 2022 Arm Limited. All rights reserved.
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#----------------------------------------------------------------------------
+#########################################################
+#            NOISE REDUCTION API library                #
+#########################################################
+cmake_minimum_required(VERSION 3.15.6)
+
+set(NOISE_REDUCTION_API_TARGET noise_reduction_api)
+project(${NOISE_REDUCTION_API_TARGET}
+        DESCRIPTION     "Noise reduction use case API library"
+        LANGUAGES       C CXX)
+
+# Create static library
+add_library(${NOISE_REDUCTION_API_TARGET} STATIC
+        src/RNNoiseProcessing.cc
+        src/RNNoiseFeatureProcessor.cc
+        src/RNNoiseModel.cc)
+
+target_include_directories(${NOISE_REDUCTION_API_TARGET} PUBLIC include)
+
+target_link_libraries(${NOISE_REDUCTION_API_TARGET} PUBLIC common_api)
+
+message(STATUS "*******************************************************")
+message(STATUS "Library                                : " ${NOISE_REDUCTION_API_TARGET})
+message(STATUS "CMAKE_SYSTEM_PROCESSOR                 : " ${CMAKE_SYSTEM_PROCESSOR})
+message(STATUS "*******************************************************")
diff --git a/source/application/api/use_case/noise_reduction/include/RNNoiseFeatureProcessor.hpp b/source/application/api/use_case/noise_reduction/include/RNNoiseFeatureProcessor.hpp
new file mode 100644
index 0000000..cbf0e4e
--- /dev/null
+++ b/source/application/api/use_case/noise_reduction/include/RNNoiseFeatureProcessor.hpp
@@ -0,0 +1,341 @@
+/*
+ * Copyright (c) 2021-2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef RNNOISE_FEATURE_PROCESSOR_HPP
+#define RNNOISE_FEATURE_PROCESSOR_HPP
+
+#include "PlatformMath.hpp"
+#include <cstdint>
+#include <vector>
+#include <array>
+#include <tuple>
+
+namespace arm {
+namespace app {
+namespace rnn {
+
+    using vec1D32F = std::vector<float>;
+    using vec2D32F = std::vector<vec1D32F>;
+    using arrHp = std::array<float, 2>;
+    using math::FftInstance;
+    using math::FftType;
+
+    class FrameFeatures {
+    public:
+        bool m_silence{false};        /* If frame contains silence or not. */
+        vec1D32F m_featuresVec{};     /* Calculated feature vector to feed to model. */
+        vec1D32F m_fftX{};            /* Vector of floats arranged to represent complex numbers. */
+        vec1D32F m_fftP{};            /* Vector of floats arranged to represent complex numbers. */
+        vec1D32F m_Ex{};              /* Spectral band energy for audio x. */
+        vec1D32F m_Ep{};              /* Spectral band energy for pitch p. */
+        vec1D32F m_Exp{};             /* Correlated spectral energy between x and p. */
+    };
+
+    /**
+     * @brief   RNNoise pre and post processing class based on the 2018 paper from
+     *          Jan-Marc Valin. Recommended reading:
+     *          - https://jmvalin.ca/demo/rnnoise/
+     *          - https://arxiv.org/abs/1709.08243
+     **/
+    class RNNoiseFeatureProcessor {
+    /* Public interface */
+    public:
+        RNNoiseFeatureProcessor();
+        ~RNNoiseFeatureProcessor() = default;
+
+        /**
+         * @brief        Calculates the features from a given audio buffer ready to be sent to RNNoise model.
+         * @param[in]    audioData   Pointer to the floating point vector
+         *                           with audio data (within the numerical
+         *                           limits of int16_t type).
+         * @param[in]    audioLen    Number of elements in the audio window.
+         * @param[out]   features    FrameFeatures object reference.
+         **/
+        void PreprocessFrame(const float*   audioData,
+                             size_t   audioLen,
+                             FrameFeatures& features);
+
+        /**
+         * @brief        Use the RNNoise model output gain values with pre-processing features
+         *               to generate audio with noise suppressed.
+         * @param[in]    modelOutput   Output gain values from model.
+         * @param[in]    features      Calculated features from pre-processing step.
+         * @param[out]   outFrame      Output frame to be populated.
+         **/
+        void PostProcessFrame(vec1D32F& modelOutput, FrameFeatures& features,  vec1D32F& outFrame);
+
+
+    /* Public constants */
+    public:
+        static constexpr uint32_t FRAME_SIZE_SHIFT{2};
+        static constexpr uint32_t FRAME_SIZE{512};
+        static constexpr uint32_t WINDOW_SIZE{2 * FRAME_SIZE};
+        static constexpr uint32_t FREQ_SIZE{FRAME_SIZE + 1};
+
+        static constexpr uint32_t PITCH_MIN_PERIOD{64};
+        static constexpr uint32_t PITCH_MAX_PERIOD{820};
+        static constexpr uint32_t PITCH_FRAME_SIZE{1024};
+        static constexpr uint32_t PITCH_BUF_SIZE{PITCH_MAX_PERIOD + PITCH_FRAME_SIZE};
+
+        static constexpr uint32_t NB_BANDS{22};
+        static constexpr uint32_t CEPS_MEM{8};
+        static constexpr uint32_t NB_DELTA_CEPS{6};
+
+        static constexpr uint32_t NB_FEATURES{NB_BANDS + 3*NB_DELTA_CEPS + 2};
+
+    /* Private functions */
+    private:
+
+        /**
+         * @brief   Initialises the half window and DCT tables.
+         */
+        void InitTables();
+
+        /**
+         * @brief           Applies a bi-quadratic filter over the audio window.
+         * @param[in]       bHp           Constant coefficient set b (arrHp type).
+         * @param[in]       aHp           Constant coefficient set a (arrHp type).
+         * @param[in,out]   memHpX        Coefficients populated by this function.
+         * @param[in,out]   audioWindow   Floating point vector with audio data.
+         **/
+        void BiQuad(
+            const arrHp& bHp,
+            const arrHp& aHp,
+            arrHp& memHpX,
+            vec1D32F& audioWindow);
+
+        /**
+         * @brief        Computes features from the "filtered" audio window.
+         * @param[in]    audioWindow   Floating point vector with audio data.
+         * @param[out]   features      FrameFeatures object reference.
+         **/
+        void ComputeFrameFeatures(vec1D32F& audioWindow, FrameFeatures& features);
+
+        /**
+         * @brief        Runs analysis on the audio buffer.
+         * @param[in]    audioWindow   Floating point vector with audio data.
+         * @param[out]   fft           Floating point FFT vector containing real and
+         *                             imaginary pairs of elements. NOTE: this vector
+         *                             does not contain the mirror image (conjugates)
+         *                             part of the spectrum.
+         * @param[out]   energy        Computed energy for each band in the Bark scale.
+         * @param[out]   analysisMem   Buffer sequentially, but partially,
+         *                             populated with new audio data.
+         **/
+        void FrameAnalysis(
+            const vec1D32F& audioWindow,
+            vec1D32F& fft,
+            vec1D32F& energy,
+            vec1D32F& analysisMem);
+
+        /**
+         * @brief               Applies the window function, in-place, over the given
+         *                      floating point buffer.
+         * @param[in,out]   x   Buffer the window will be applied to.
+         **/
+        void ApplyWindow(vec1D32F& x);
+
+        /**
+         * @brief        Computes the FFT for a given vector.
+         * @param[in]    x     Vector to compute the FFT from.
+         * @param[out]   fft   Floating point FFT vector containing real and
+         *                     imaginary pairs of elements. NOTE: this vector
+         *                     does not contain the mirror image (conjugates)
+         *                     part of the spectrum.
+         **/
+        void ForwardTransform(
+            vec1D32F& x,
+            vec1D32F& fft);
+
+        /**
+         * @brief        Computes band energy for each of the 22 Bark scale bands.
+         * @param[in]    fft_X   FFT spectrum (as computed by ForwardTransform).
+         * @param[out]   bandE   Vector with 22 elements populated with energy for
+         *                       each band.
+         **/
+        void ComputeBandEnergy(const vec1D32F& fft_X, vec1D32F& bandE);
+
+        /**
+         * @brief        Computes band energy correlation.
+         * @param[in]    X       FFT vector X.
+         * @param[in]    P       FFT vector P.
+         * @param[out]   bandC   Vector with 22 elements populated with band energy
+         *                       correlation for the two input FFT vectors.
+         **/
+        void ComputeBandCorr(const vec1D32F& X, const vec1D32F& P, vec1D32F& bandC);
+
+        /**
+         * @brief        Performs pitch auto-correlation for a given vector for
+         *               given lag.
+         * @param[in]    x     Input vector.
+         * @param[out]   ac    Auto-correlation output vector.
+         * @param[in]    lag   Lag value.
+         * @param[in]    n     Number of elements to consider for correlation
+         *                     computation.
+         **/
+        void AutoCorr(const vec1D32F &x,
+                     vec1D32F &ac,
+                     size_t lag,
+                     size_t n);
+
+        /**
+         * @brief       Computes pitch cross-correlation.
+         * @param[in]   x          Input vector 1.
+         * @param[in]   y          Input vector 2.
+         * @param[out]  xCorr         Cross-correlation output vector.
+         * @param[in]   len        Number of elements to consider for correlation.
+         *                         computation.
+         * @param[in]   maxPitch   Maximum pitch.
+         **/
+        void PitchXCorr(
+            const vec1D32F& x,
+            const vec1D32F& y,
+            vec1D32F& xCorr,
+            size_t len,
+            size_t maxPitch);
+
+        /**
+         * @brief        Computes "Linear Predictor Coefficients".
+         * @param[in]    ac    Correlation vector.
+         * @param[in]    p     Number of elements of input vector to consider.
+         * @param[out]   lpc   Output coefficients vector.
+         **/
+        void LPC(const vec1D32F& ac, int32_t p, vec1D32F& lpc);
+
+        /**
+         * @brief        Custom FIR implementation.
+         * @param[in]    num   FIR coefficient vector.
+         * @param[in]    N     Number of elements.
+         * @param[out]   x     Vector to be be processed.
+         **/
+        void Fir5(const vec1D32F& num, uint32_t N, vec1D32F& x);
+
+        /**
+         * @brief           Down-sample the pitch buffer.
+         * @param[in,out]   pitchBuf     Pitch buffer.
+         * @param[in]       pitchBufSz   Buffer size.
+         **/
+        void PitchDownsample(vec1D32F& pitchBuf, size_t pitchBufSz);
+
+        /**
+         * @brief       Pitch search function.
+         * @param[in]   xLP        Shifted pitch buffer input.
+         * @param[in]   y          Pitch buffer input.
+         * @param[in]   len        Length to search for.
+         * @param[in]   maxPitch   Maximum pitch.
+         * @return      pitch index.
+         **/
+        int PitchSearch(vec1D32F& xLp, vec1D32F& y, uint32_t len, uint32_t maxPitch);
+
+        /**
+         * @brief       Finds the "best" pitch from the buffer.
+         * @param[in]   xCorr      Pitch correlation vector.
+         * @param[in]   y          Pitch buffer input.
+         * @param[in]   len        Length to search for.
+         * @param[in]   maxPitch   Maximum pitch.
+         * @return      pitch array (2 elements).
+         **/
+        arrHp FindBestPitch(vec1D32F& xCorr, vec1D32F& y, uint32_t len, uint32_t maxPitch);
+
+        /**
+         * @brief           Remove pitch period doubling errors.
+         * @param[in,out]   pitchBuf     Pitch buffer vector.
+         * @param[in]       maxPeriod    Maximum period.
+         * @param[in]       minPeriod    Minimum period.
+         * @param[in]       frameSize    Frame size.
+         * @param[in]       pitchIdx0_   Pitch index 0.
+         * @return          pitch index.
+         **/
+        int RemoveDoubling(
+                vec1D32F& pitchBuf,
+                uint32_t maxPeriod,
+                uint32_t minPeriod,
+                uint32_t frameSize,
+                size_t pitchIdx0_);
+
+        /**
+         * @brief       Computes pitch gain.
+         * @param[in]   xy   Single xy cross correlation value.
+         * @param[in]   xx   Single xx auto correlation value.
+         * @param[in]   yy   Single yy auto correlation value.
+         * @return      Calculated pitch gain.
+         **/
+        float ComputePitchGain(float xy, float xx, float yy);
+
+        /**
+         * @brief        Computes DCT vector from the given input.
+         * @param[in]    input    Input vector.
+         * @param[out]   output   Output vector with DCT coefficients.
+         **/
+        void DCT(vec1D32F& input, vec1D32F& output);
+
+        /**
+         * @brief        Perform inverse fourier transform on complex spectral vector.
+         * @param[out]   out      Output vector.
+         * @param[in]    fftXIn   Vector of floats arranged to represent complex numbers interleaved.
+         **/
+        void InverseTransform(vec1D32F& out, vec1D32F& fftXIn);
+
+        /**
+         * @brief       Perform pitch filtering.
+         * @param[in]   features   Object with pre-processing calculated frame features.
+         * @param[in]   g          Gain values.
+         **/
+        void PitchFilter(FrameFeatures& features, vec1D32F& g);
+
+        /**
+         * @brief        Interpolate the band gain values.
+         * @param[out]   g       Gain values.
+         * @param[in]    bandE   Vector with 22 elements populated with energy for
+         *                       each band.
+         **/
+        void InterpBandGain(vec1D32F& g, vec1D32F& bandE);
+
+        /**
+         * @brief        Create de-noised frame.
+         * @param[out]   outFrame   Output vector for storing the created audio frame.
+         * @param[in]    fftY       Gain adjusted complex spectral vector.
+         */
+        void FrameSynthesis(vec1D32F& outFrame, vec1D32F& fftY);
+
+    /* Private objects */
+    private:
+        FftInstance m_fftInstReal;  /* FFT instance for real numbers */
+        FftInstance m_fftInstCmplx; /* FFT instance for complex numbers */
+        vec1D32F m_halfWindow;      /* Window coefficients */
+        vec1D32F m_dctTable;        /* DCT table */
+        vec1D32F m_analysisMem;     /* Buffer used for frame analysis */
+        vec2D32F m_cepstralMem;     /* Cepstral coefficients */
+        size_t m_memId;             /* memory ID */
+        vec1D32F m_synthesisMem;    /* Synthesis mem (used by post-processing) */
+        vec1D32F m_pitchBuf;        /* Pitch buffer */
+        float m_lastGain;           /* Last gain calculated */
+        int m_lastPeriod;           /* Last period calculated */
+        arrHp m_memHpX;             /* HpX coefficients. */
+        vec1D32F m_lastGVec;        /* Last gain vector (used by post-processing) */
+
+        /* Constants */
+        const std::array <uint32_t, NB_BANDS> m_eband5ms {
+            0,  1,  2,  3,  4,  5,  6,  7,  8, 10,  12,
+            14, 16, 20, 24, 28, 34, 40, 48, 60, 78, 100};
+    };
+
+
+} /* namespace rnn */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* RNNOISE_FEATURE_PROCESSOR_HPP */
diff --git a/source/application/api/use_case/noise_reduction/include/RNNoiseModel.hpp b/source/application/api/use_case/noise_reduction/include/RNNoiseModel.hpp
new file mode 100644
index 0000000..3d2f23c
--- /dev/null
+++ b/source/application/api/use_case/noise_reduction/include/RNNoiseModel.hpp
@@ -0,0 +1,78 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef RNNOISE_MODEL_HPP
+#define RNNOISE_MODEL_HPP
+
+#include "Model.hpp"
+
+extern const uint32_t g_NumInputFeatures;
+extern const uint32_t g_FrameLength;
+extern const uint32_t g_FrameStride;
+
+namespace arm {
+namespace app {
+
+    class RNNoiseModel : public Model {
+    public:
+        /**
+         * @brief Runs inference for RNNoise model.
+         *
+         * Call CopyGruStates so GRU state outputs are copied to GRU state inputs before the inference run.
+         * Run ResetGruState() method to set states to zero before starting processing logically related data.
+         * @return True if inference succeeded, False - otherwise
+         */
+        bool RunInference() override;
+
+        /**
+         * @brief Sets GRU input states to zeros.
+         * Call this method before starting processing the new sequence of logically related data.
+         */
+        void ResetGruState();
+
+        /**
+        * @brief Copy current GRU output states to input states.
+        * Call this method before starting processing the next sequence of logically related data.
+         */
+        bool CopyGruStates();
+
+        /* Which index of model outputs does the main output (gains) come from. */
+        const size_t m_indexForModelOutput = 1;
+
+    protected:
+        /** @brief   Gets the reference to op resolver interface class. */
+        const tflite::MicroOpResolver& GetOpResolver() override;
+
+        /** @brief   Adds operations to the op resolver instance. */
+        bool EnlistOperations() override;
+
+        /*
+        Each inference after the first needs to copy 3 GRU states from a output index to input index (model dependent):
+        0 -> 3, 2 -> 2, 3 -> 1
+        */
+        const std::vector<std::pair<size_t, size_t>> m_gruStateMap = {{0,3}, {2, 2}, {3, 1}};
+    private:
+        /* Maximum number of individual operations that can be enlisted. */
+        static constexpr int ms_maxOpCnt = 15;
+
+        /* A mutable op resolver instance. */
+        tflite::MicroMutableOpResolver<ms_maxOpCnt> m_opResolver;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* RNNOISE_MODEL_HPP */
diff --git a/source/application/api/use_case/noise_reduction/include/RNNoiseProcessing.hpp b/source/application/api/use_case/noise_reduction/include/RNNoiseProcessing.hpp
new file mode 100644
index 0000000..15e62d9
--- /dev/null
+++ b/source/application/api/use_case/noise_reduction/include/RNNoiseProcessing.hpp
@@ -0,0 +1,113 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef RNNOISE_PROCESSING_HPP
+#define RNNOISE_PROCESSING_HPP
+
+#include "BaseProcessing.hpp"
+#include "Model.hpp"
+#include "RNNoiseFeatureProcessor.hpp"
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   Pre-processing class for Noise Reduction use case.
+     *          Implements methods declared by BasePreProcess and anything else needed
+     *          to populate input tensors ready for inference.
+     */
+    class RNNoisePreProcess : public BasePreProcess {
+
+    public:
+        /**
+         * @brief           Constructor
+         * @param[in]       inputTensor        Pointer to the TFLite Micro input Tensor.
+         * @param[in/out]   featureProcessor   RNNoise specific feature extractor object.
+         * @param[in/out]   frameFeatures      RNNoise specific features shared between pre & post-processing.
+         *
+         **/
+        explicit RNNoisePreProcess(TfLiteTensor* inputTensor,
+                                   std::shared_ptr<rnn::RNNoiseFeatureProcessor> featureProcessor,
+                                   std::shared_ptr<rnn::FrameFeatures> frameFeatures);
+
+        /**
+         * @brief       Should perform pre-processing of 'raw' input audio data and load it into
+         *              TFLite Micro input tensors ready for inference
+         * @param[in]   input      Pointer to the data that pre-processing will work on.
+         * @param[in]   inputSize  Size of the input data.
+         * @return      true if successful, false otherwise.
+         **/
+        bool DoPreProcess(const void* input, size_t inputSize) override;
+
+    private:
+        TfLiteTensor* m_inputTensor;                        /* Model input tensor. */
+        std::shared_ptr<rnn::RNNoiseFeatureProcessor> m_featureProcessor;   /* RNNoise feature processor shared between pre & post-processing. */
+        std::shared_ptr<rnn::FrameFeatures> m_frameFeatures;                /* RNNoise features shared between pre & post-processing. */
+        rnn::vec1D32F m_audioFrame;                         /* Audio frame cast to FP32 */
+
+        /**
+         * @brief            Quantize the given features and populate the input Tensor.
+         * @param[in]        inputFeatures   Vector of floating point features to quantize.
+         * @param[in]        quantScale      Quantization scale for the inputTensor.
+         * @param[in]        quantOffset     Quantization offset for the inputTensor.
+         * @param[in,out]    inputTensor     TFLite micro tensor to populate.
+         **/
+        static void QuantizeAndPopulateInput(rnn::vec1D32F& inputFeatures,
+                float quantScale, int quantOffset,
+                TfLiteTensor* inputTensor);
+    };
+
+    /**
+     * @brief   Post-processing class for Noise Reduction use case.
+     *          Implements methods declared by BasePostProcess and anything else needed
+     *          to populate result vector.
+     */
+    class RNNoisePostProcess : public BasePostProcess {
+
+    public:
+        /**
+         * @brief           Constructor
+         * @param[in]       outputTensor         Pointer to the TFLite Micro output Tensor.
+         * @param[out]      denoisedAudioFrame   Vector to store the final denoised audio frame.
+         * @param[in/out]   featureProcessor     RNNoise specific feature extractor object.
+         * @param[in/out]   frameFeatures        RNNoise specific features shared between pre & post-processing.
+         **/
+        RNNoisePostProcess(TfLiteTensor* outputTensor,
+                           std::vector<int16_t>& denoisedAudioFrame,
+                           std::shared_ptr<rnn::RNNoiseFeatureProcessor> featureProcessor,
+                           std::shared_ptr<rnn::FrameFeatures> frameFeatures);
+
+        /**
+         * @brief       Should perform post-processing of the result of inference then
+         *              populate result data for any later use.
+         * @return      true if successful, false otherwise.
+         **/
+        bool DoPostProcess() override;
+
+    private:
+        TfLiteTensor* m_outputTensor;                       /* Model output tensor. */
+        std::vector<int16_t>& m_denoisedAudioFrame;         /* Vector to store the final denoised frame. */
+        rnn::vec1D32F m_denoisedAudioFrameFloat;            /* Internal vector to store the final denoised frame (FP32). */
+        std::shared_ptr<rnn::RNNoiseFeatureProcessor> m_featureProcessor;   /* RNNoise feature processor shared between pre & post-processing. */
+        std::shared_ptr<rnn::FrameFeatures> m_frameFeatures;                /* RNNoise features shared between pre & post-processing. */
+        std::vector<float> m_modelOutputFloat;              /* Internal vector to store de-quantized model output. */
+
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* RNNOISE_PROCESSING_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/noise_reduction/src/RNNoiseFeatureProcessor.cc b/source/application/api/use_case/noise_reduction/src/RNNoiseFeatureProcessor.cc
new file mode 100644
index 0000000..036894c
--- /dev/null
+++ b/source/application/api/use_case/noise_reduction/src/RNNoiseFeatureProcessor.cc
@@ -0,0 +1,892 @@
+/*
+ * Copyright (c) 2021-2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "RNNoiseFeatureProcessor.hpp"
+#include "log_macros.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstring>
+
+namespace arm {
+namespace app {
+namespace rnn {
+
+#define VERIFY(x)                                   \
+do {                                                \
+    if (!(x)) {                                     \
+        printf_err("Assert failed:" #x "\n");       \
+        exit(1);                                    \
+    }                                               \
+} while(0)
+
+RNNoiseFeatureProcessor::RNNoiseFeatureProcessor() :
+        m_halfWindow(FRAME_SIZE, 0),
+        m_dctTable(NB_BANDS * NB_BANDS),
+        m_analysisMem(FRAME_SIZE, 0),
+        m_cepstralMem(CEPS_MEM, vec1D32F(NB_BANDS, 0)),
+        m_memId{0},
+        m_synthesisMem(FRAME_SIZE, 0),
+        m_pitchBuf(PITCH_BUF_SIZE, 0),
+        m_lastGain{0.0},
+        m_lastPeriod{0},
+        m_memHpX{},
+        m_lastGVec(NB_BANDS, 0)
+{
+    constexpr uint32_t numFFt = 2 * FRAME_SIZE;
+    static_assert(numFFt != 0, "Num FFT can't be 0");
+
+    math::MathUtils::FftInitF32(numFFt, this->m_fftInstReal, FftType::real);
+    math::MathUtils::FftInitF32(numFFt, this->m_fftInstCmplx, FftType::complex);
+    this->InitTables();
+}
+
+void RNNoiseFeatureProcessor::PreprocessFrame(const float*   audioData,
+                                              const size_t   audioLen,
+                                              FrameFeatures& features)
+{
+    /* Note audioWindow is modified in place */
+    const arrHp aHp {-1.99599, 0.99600 };
+    const arrHp bHp {-2.00000, 1.00000 };
+
+    vec1D32F audioWindow{audioData, audioData + audioLen};
+
+    this->BiQuad(bHp, aHp, this->m_memHpX, audioWindow);
+    this->ComputeFrameFeatures(audioWindow, features);
+}
+
+void RNNoiseFeatureProcessor::PostProcessFrame(vec1D32F& modelOutput, FrameFeatures& features, vec1D32F& outFrame)
+{
+    std::vector<float> outputBands = modelOutput;
+    std::vector<float> gain(FREQ_SIZE, 0);
+
+    if (!features.m_silence) {
+        PitchFilter(features, outputBands);
+        for (size_t i = 0; i < NB_BANDS; i++) {
+            float alpha = .6f;
+            outputBands[i] = std::max(outputBands[i], alpha * m_lastGVec[i]);
+            m_lastGVec[i] = outputBands[i];
+        }
+        InterpBandGain(gain, outputBands);
+        for (size_t i = 0; i < FREQ_SIZE; i++) {
+            features.m_fftX[2 * i] *= gain[i];  /* Real. */
+            features.m_fftX[2 * i + 1] *= gain[i];  /*imaginary. */
+
+        }
+
+    }
+
+    FrameSynthesis(outFrame, features.m_fftX);
+}
+
+void RNNoiseFeatureProcessor::InitTables()
+{
+    constexpr float pi = M_PI;
+    constexpr float halfPi = M_PI / 2;
+    constexpr float halfPiOverFrameSz = halfPi/FRAME_SIZE;
+
+    for (uint32_t i = 0; i < FRAME_SIZE; i++) {
+        const float sinVal = math::MathUtils::SineF32(halfPiOverFrameSz * (i + 0.5f));
+        m_halfWindow[i] = math::MathUtils::SineF32(halfPi * sinVal * sinVal);
+    }
+
+    for (uint32_t i = 0; i < NB_BANDS; i++) {
+        for (uint32_t j = 0; j < NB_BANDS; j++) {
+            m_dctTable[i * NB_BANDS + j] = math::MathUtils::CosineF32((i + 0.5f) * j * pi / NB_BANDS);
+        }
+        m_dctTable[i * NB_BANDS] *= math::MathUtils::SqrtF32(0.5f);
+    }
+}
+
+void RNNoiseFeatureProcessor::BiQuad(
+        const arrHp& bHp,
+        const arrHp& aHp,
+        arrHp& memHpX,
+        vec1D32F& audioWindow)
+{
+    for (float& audioElement : audioWindow) {
+        const auto xi = audioElement;
+        const auto yi = audioElement + memHpX[0];
+        memHpX[0] = memHpX[1] + (bHp[0] * xi - aHp[0] * yi);
+        memHpX[1] = (bHp[1] * xi - aHp[1] * yi);
+        audioElement = yi;
+    }
+}
+
+void RNNoiseFeatureProcessor::ComputeFrameFeatures(vec1D32F& audioWindow,
+                                                   FrameFeatures& features)
+{
+    this->FrameAnalysis(audioWindow,
+                        features.m_fftX,
+                        features.m_Ex,
+                        this->m_analysisMem);
+
+    float energy = 0.0;
+
+    vec1D32F Ly(NB_BANDS, 0);
+    vec1D32F p(WINDOW_SIZE, 0);
+    vec1D32F pitchBuf(PITCH_BUF_SIZE >> 1, 0);
+
+    VERIFY(PITCH_BUF_SIZE >= this->m_pitchBuf.size());
+    std::copy_n(this->m_pitchBuf.begin() + FRAME_SIZE,
+                PITCH_BUF_SIZE - FRAME_SIZE,
+                this->m_pitchBuf.begin());
+
+    VERIFY(FRAME_SIZE <= audioWindow.size() && PITCH_BUF_SIZE > FRAME_SIZE);
+    std::copy_n(audioWindow.begin(),
+                FRAME_SIZE,
+                this->m_pitchBuf.begin() + PITCH_BUF_SIZE - FRAME_SIZE);
+
+    this->PitchDownsample(pitchBuf, PITCH_BUF_SIZE);
+
+    VERIFY(pitchBuf.size() > PITCH_MAX_PERIOD/2);
+    vec1D32F xLp(pitchBuf.size() - PITCH_MAX_PERIOD/2, 0);
+    std::copy_n(pitchBuf.begin() + PITCH_MAX_PERIOD/2, xLp.size(), xLp.begin());
+
+    int pitchIdx = this->PitchSearch(xLp, pitchBuf,
+            PITCH_FRAME_SIZE, (PITCH_MAX_PERIOD - (3*PITCH_MIN_PERIOD)));
+
+    pitchIdx = this->RemoveDoubling(
+                pitchBuf,
+                PITCH_MAX_PERIOD,
+                PITCH_MIN_PERIOD,
+                PITCH_FRAME_SIZE,
+                PITCH_MAX_PERIOD - pitchIdx);
+
+    size_t stIdx = PITCH_BUF_SIZE - WINDOW_SIZE - pitchIdx;
+    VERIFY((static_cast<int>(PITCH_BUF_SIZE) - static_cast<int>(WINDOW_SIZE) - pitchIdx) >= 0);
+    std::copy_n(this->m_pitchBuf.begin() + stIdx, WINDOW_SIZE, p.begin());
+
+    this->ApplyWindow(p);
+    this->ForwardTransform(p, features.m_fftP);
+    this->ComputeBandEnergy(features.m_fftP, features.m_Ep);
+    this->ComputeBandCorr(features.m_fftX, features.m_fftP, features.m_Exp);
+
+    for (uint32_t i = 0 ; i < NB_BANDS; ++i) {
+        features.m_Exp[i] /= math::MathUtils::SqrtF32(
+            0.001f + features.m_Ex[i] * features.m_Ep[i]);
+    }
+
+    vec1D32F dctVec(NB_BANDS, 0);
+    this->DCT(features.m_Exp, dctVec);
+
+    features.m_featuresVec = vec1D32F (NB_FEATURES, 0);
+    for (uint32_t i = 0; i < NB_DELTA_CEPS; ++i) {
+        features.m_featuresVec[NB_BANDS + 2*NB_DELTA_CEPS + i] = dctVec[i];
+    }
+
+    features.m_featuresVec[NB_BANDS + 2*NB_DELTA_CEPS] -= 1.3;
+    features.m_featuresVec[NB_BANDS + 2*NB_DELTA_CEPS + 1] -= 0.9;
+    features.m_featuresVec[NB_BANDS + 3*NB_DELTA_CEPS] = 0.01 * (static_cast<int>(pitchIdx) - 300);
+
+    float logMax = -2.f;
+    float follow = -2.f;
+    for (uint32_t i = 0; i < NB_BANDS; ++i) {
+        Ly[i] = log10f(1e-2f + features.m_Ex[i]);
+        Ly[i] = std::max<float>(logMax - 7, std::max<float>(follow - 1.5, Ly[i]));
+        logMax = std::max<float>(logMax, Ly[i]);
+        follow = std::max<float>(follow - 1.5, Ly[i]);
+        energy += features.m_Ex[i];
+    }
+
+    /* If there's no audio avoid messing up the state. */
+    features.m_silence = true;
+    if (energy < 0.04) {
+        return;
+    } else {
+        features.m_silence = false;
+    }
+
+    this->DCT(Ly, features.m_featuresVec);
+    features.m_featuresVec[0] -= 12.0;
+    features.m_featuresVec[1] -= 4.0;
+
+    VERIFY(CEPS_MEM > 2);
+    uint32_t stIdx1 = this->m_memId < 1 ? CEPS_MEM + this->m_memId - 1 : this->m_memId - 1;
+    uint32_t stIdx2 = this->m_memId < 2 ? CEPS_MEM + this->m_memId - 2 : this->m_memId - 2;
+    VERIFY(stIdx1 < this->m_cepstralMem.size());
+    VERIFY(stIdx2 < this->m_cepstralMem.size());
+    auto ceps1 = this->m_cepstralMem[stIdx1];
+    auto ceps2 = this->m_cepstralMem[stIdx2];
+
+    /* Ceps 0 */
+    for (uint32_t i = 0; i < NB_BANDS; ++i) {
+        this->m_cepstralMem[this->m_memId][i] = features.m_featuresVec[i];
+    }
+
+    for (uint32_t i = 0; i < NB_DELTA_CEPS; ++i) {
+        features.m_featuresVec[i] = this->m_cepstralMem[this->m_memId][i] + ceps1[i] + ceps2[i];
+        features.m_featuresVec[NB_BANDS + i] = this->m_cepstralMem[this->m_memId][i] - ceps2[i];
+        features.m_featuresVec[NB_BANDS + NB_DELTA_CEPS + i] =
+                this->m_cepstralMem[this->m_memId][i] - 2 * ceps1[i] + ceps2[i];
+    }
+
+    /* Spectral variability features. */
+    this->m_memId += 1;
+    if (this->m_memId == CEPS_MEM) {
+        this->m_memId = 0;
+    }
+
+    float specVariability = 0.f;
+
+    VERIFY(this->m_cepstralMem.size() >= CEPS_MEM);
+    for (size_t i = 0; i < CEPS_MEM; ++i) {
+        float minDist = 1e15;
+        for (size_t j = 0; j < CEPS_MEM; ++j) {
+            float dist = 0.f;
+            for (size_t k = 0; k < NB_BANDS; ++k) {
+                VERIFY(this->m_cepstralMem[i].size() >= NB_BANDS);
+                auto tmp = this->m_cepstralMem[i][k] - this->m_cepstralMem[j][k];
+                dist += tmp * tmp;
+            }
+
+            if (j != i) {
+                minDist = std::min<float>(minDist, dist);
+            }
+        }
+        specVariability += minDist;
+    }
+
+    VERIFY(features.m_featuresVec.size() >= NB_BANDS + 3 * NB_DELTA_CEPS + 1);
+    features.m_featuresVec[NB_BANDS + 3 * NB_DELTA_CEPS + 1] = specVariability / CEPS_MEM - 2.1;
+}
+
+void RNNoiseFeatureProcessor::FrameAnalysis(
+    const vec1D32F& audioWindow,
+    vec1D32F& fft,
+    vec1D32F& energy,
+    vec1D32F& analysisMem)
+{
+    vec1D32F x(WINDOW_SIZE, 0);
+
+    /* Move old audio down and populate end with latest audio window. */
+    VERIFY(x.size() >= FRAME_SIZE && analysisMem.size() >= FRAME_SIZE);
+    VERIFY(audioWindow.size() >= FRAME_SIZE);
+
+    std::copy_n(analysisMem.begin(), FRAME_SIZE, x.begin());
+    std::copy_n(audioWindow.begin(), x.size() - FRAME_SIZE, x.begin() + FRAME_SIZE);
+    std::copy_n(audioWindow.begin(), FRAME_SIZE, analysisMem.begin());
+
+    this->ApplyWindow(x);
+
+    /* Calculate FFT. */
+    ForwardTransform(x, fft);
+
+    /* Compute band energy. */
+    ComputeBandEnergy(fft, energy);
+}
+
+void RNNoiseFeatureProcessor::ApplyWindow(vec1D32F& x)
+{
+    if (WINDOW_SIZE != x.size()) {
+        printf_err("Invalid size for vector to be windowed\n");
+        return;
+    }
+
+    VERIFY(this->m_halfWindow.size() >= FRAME_SIZE);
+
+    /* Multiply input by sinusoidal function. */
+    for (size_t i = 0; i < FRAME_SIZE; i++) {
+        x[i] *= this->m_halfWindow[i];
+        x[WINDOW_SIZE - 1 - i] *= this->m_halfWindow[i];
+    }
+}
+
+void RNNoiseFeatureProcessor::ForwardTransform(
+    vec1D32F& x,
+    vec1D32F& fft)
+{
+    /* The input vector can be modified by the fft function. */
+    fft.reserve(x.size() + 2);
+    fft.resize(x.size() + 2, 0);
+    math::MathUtils::FftF32(x, fft, this->m_fftInstReal);
+
+    /* Normalise. */
+    for (auto& f : fft) {
+        f /= this->m_fftInstReal.m_fftLen;
+    }
+
+    /* Place the last freq element correctly */
+    fft[fft.size()-2] = fft[1];
+    fft[1] = 0;
+
+    /* NOTE: We don't truncate out FFT vector as it already contains only the
+     * first half of the FFT's. The conjugates are not present. */
+}
+
+void RNNoiseFeatureProcessor::ComputeBandEnergy(const vec1D32F& fftX, vec1D32F& bandE)
+{
+    bandE = vec1D32F(NB_BANDS, 0);
+
+    VERIFY(this->m_eband5ms.size() >= NB_BANDS);
+    for (uint32_t i = 0; i < NB_BANDS - 1; i++) {
+        const auto bandSize = (this->m_eband5ms[i + 1] - this->m_eband5ms[i])
+                              << FRAME_SIZE_SHIFT;
+
+        for (uint32_t j = 0; j < bandSize; j++) {
+            const auto frac = static_cast<float>(j) / bandSize;
+            const auto idx = (this->m_eband5ms[i] << FRAME_SIZE_SHIFT) + j;
+
+            auto tmp = fftX[2 * idx] * fftX[2 * idx]; /* Real part */
+            tmp += fftX[2 * idx + 1] * fftX[2 * idx + 1]; /* Imaginary part */
+
+            bandE[i] += (1 - frac) * tmp;
+            bandE[i + 1] += frac * tmp;
+        }
+    }
+    bandE[0] *= 2;
+    bandE[NB_BANDS - 1] *= 2;
+}
+
+void RNNoiseFeatureProcessor::ComputeBandCorr(const vec1D32F& X, const vec1D32F& P, vec1D32F& bandC)
+{
+    bandC = vec1D32F(NB_BANDS, 0);
+    VERIFY(this->m_eband5ms.size() >= NB_BANDS);
+
+    for (uint32_t i = 0; i < NB_BANDS - 1; i++) {
+        const auto bandSize = (this->m_eband5ms[i + 1] - this->m_eband5ms[i]) << FRAME_SIZE_SHIFT;
+
+        for (uint32_t j = 0; j < bandSize; j++) {
+            const auto frac = static_cast<float>(j) / bandSize;
+            const auto idx = (this->m_eband5ms[i] << FRAME_SIZE_SHIFT) + j;
+
+            auto tmp = X[2 * idx] * P[2 * idx]; /* Real part */
+            tmp += X[2 * idx + 1] * P[2 * idx + 1]; /* Imaginary part */
+
+            bandC[i] += (1 - frac) * tmp;
+            bandC[i + 1] += frac * tmp;
+        }
+    }
+    bandC[0] *= 2;
+    bandC[NB_BANDS - 1] *= 2;
+}
+
+void RNNoiseFeatureProcessor::DCT(vec1D32F& input, vec1D32F& output)
+{
+    VERIFY(this->m_dctTable.size() >= NB_BANDS * NB_BANDS);
+    for (uint32_t i = 0; i < NB_BANDS; ++i) {
+        float sum = 0;
+
+        for (uint32_t j = 0, k = 0; j < NB_BANDS; ++j, k += NB_BANDS) {
+            sum += input[j] * this->m_dctTable[k + i];
+        }
+        output[i] = sum * math::MathUtils::SqrtF32(2.0/22);
+    }
+}
+
+void RNNoiseFeatureProcessor::PitchDownsample(vec1D32F& pitchBuf, size_t pitchBufSz) {
+    for (size_t i = 1; i < (pitchBufSz >> 1); ++i) {
+        pitchBuf[i] = 0.5 * (
+                        0.5 * (this->m_pitchBuf[2 * i - 1] + this->m_pitchBuf[2 * i + 1])
+                            + this->m_pitchBuf[2 * i]);
+    }
+
+    pitchBuf[0] = 0.5*(0.5*(this->m_pitchBuf[1]) + this->m_pitchBuf[0]);
+
+    vec1D32F ac(5, 0);
+    size_t numLags = 4;
+
+    this->AutoCorr(pitchBuf, ac, numLags, pitchBufSz >> 1);
+
+    /* Noise floor -40db */
+    ac[0] *= 1.0001;
+
+    /* Lag windowing. */
+    for (size_t i = 1; i < numLags + 1; ++i) {
+        ac[i] -= ac[i] * (0.008 * i) * (0.008 * i);
+    }
+
+    vec1D32F lpc(numLags, 0);
+    this->LPC(ac, numLags, lpc);
+
+    float tmp = 1.0;
+    for (size_t i = 0; i < numLags; ++i) {
+        tmp = 0.9f * tmp;
+        lpc[i] = lpc[i] * tmp;
+    }
+
+    vec1D32F lpc2(numLags + 1, 0);
+    float c1 = 0.8;
+
+    /* Add a zero. */
+    lpc2[0] = lpc[0] + 0.8;
+    lpc2[1] = lpc[1] + (c1 * lpc[0]);
+    lpc2[2] = lpc[2] + (c1 * lpc[1]);
+    lpc2[3] = lpc[3] + (c1 * lpc[2]);
+    lpc2[4] = (c1 * lpc[3]);
+
+    this->Fir5(lpc2, pitchBufSz >> 1, pitchBuf);
+}
+
+int RNNoiseFeatureProcessor::PitchSearch(vec1D32F& xLp, vec1D32F& y, uint32_t len, uint32_t maxPitch) {
+    uint32_t lag = len + maxPitch;
+    vec1D32F xLp4(len >> 2, 0);
+    vec1D32F yLp4(lag >> 2, 0);
+    vec1D32F xCorr(maxPitch >> 1, 0);
+
+    /* Downsample by 2 again. */
+    for (size_t j = 0; j < (len >> 2); ++j) {
+        xLp4[j] = xLp[2*j];
+    }
+    for (size_t j = 0; j < (lag >> 2); ++j) {
+        yLp4[j] = y[2*j];
+    }
+
+    this->PitchXCorr(xLp4, yLp4, xCorr, len >> 2, maxPitch >> 2);
+
+    /* Coarse search with 4x decimation. */
+    arrHp bestPitch = this->FindBestPitch(xCorr, yLp4, len >> 2, maxPitch >> 2);
+
+    /* Finer search with 2x decimation. */
+    const int maxIdx = (maxPitch >> 1);
+    for (int i = 0; i < maxIdx; ++i) {
+        xCorr[i] = 0;
+        if (std::abs(i - 2*bestPitch[0]) > 2 and std::abs(i - 2*bestPitch[1]) > 2) {
+            continue;
+        }
+        float sum = 0;
+        for (size_t j = 0; j < len >> 1; ++j) {
+            sum += xLp[j] * y[i+j];
+        }
+
+        xCorr[i] = std::max(-1.0f, sum);
+    }
+
+    bestPitch = this->FindBestPitch(xCorr, y, len >> 1, maxPitch >> 1);
+
+    int offset;
+    /* Refine by pseudo-interpolation. */
+    if ( 0 < bestPitch[0] && bestPitch[0] < ((maxPitch >> 1) - 1)) {
+        float a = xCorr[bestPitch[0] - 1];
+        float b = xCorr[bestPitch[0]];
+        float c = xCorr[bestPitch[0] + 1];
+
+        if ( (c-a) > 0.7*(b-a) ) {
+            offset = 1;
+        } else if ( (a-c) > 0.7*(b-c) ) {
+            offset = -1;
+        } else {
+            offset = 0;
+        }
+    } else {
+        offset = 0;
+    }
+
+    return 2*bestPitch[0] - offset;
+}
+
+arrHp RNNoiseFeatureProcessor::FindBestPitch(vec1D32F& xCorr, vec1D32F& y, uint32_t len, uint32_t maxPitch)
+{
+    float Syy = 1;
+    arrHp bestNum {-1, -1};
+    arrHp bestDen {0, 0};
+    arrHp bestPitch {0, 1};
+
+    for (size_t j = 0; j < len; ++j) {
+        Syy += (y[j] * y[j]);
+    }
+
+    for (size_t i = 0; i < maxPitch; ++i ) {
+        if (xCorr[i] > 0) {
+            float xCorr16 = xCorr[i] * 1e-12f;  /* Avoid problems when squaring. */
+
+            float num = xCorr16 * xCorr16;
+            if (num*bestDen[1] > bestNum[1]*Syy) {
+                if (num*bestDen[0] > bestNum[0]*Syy) {
+                    bestNum[1] = bestNum[0];
+                    bestDen[1] = bestDen[0];
+                    bestPitch[1] = bestPitch[0];
+                    bestNum[0] = num;
+                    bestDen[0] = Syy;
+                    bestPitch[0] = i;
+                } else {
+                    bestNum[1] = num;
+                    bestDen[1] = Syy;
+                    bestPitch[1] = i;
+                }
+            }
+        }
+
+        Syy += (y[i+len]*y[i+len]) - (y[i]*y[i]);
+        Syy = std::max(1.0f, Syy);
+    }
+
+    return bestPitch;
+}
+
+int RNNoiseFeatureProcessor::RemoveDoubling(
+    vec1D32F& pitchBuf,
+    uint32_t maxPeriod,
+    uint32_t minPeriod,
+    uint32_t frameSize,
+    size_t pitchIdx0_)
+{
+    constexpr std::array<size_t, 16> secondCheck {0, 0, 3, 2, 3, 2, 5, 2, 3, 2, 3, 2, 5, 2, 3, 2};
+    uint32_t minPeriod0 = minPeriod;
+    float lastPeriod = static_cast<float>(this->m_lastPeriod)/2;
+    float lastGain = static_cast<float>(this->m_lastGain);
+
+    maxPeriod /= 2;
+    minPeriod /= 2;
+    pitchIdx0_ /= 2;
+    frameSize /= 2;
+    uint32_t xStart = maxPeriod;
+
+    if (pitchIdx0_ >= maxPeriod) {
+        pitchIdx0_ = maxPeriod - 1;
+    }
+
+    size_t pitchIdx  = pitchIdx0_;
+    const size_t pitchIdx0 = pitchIdx0_;
+
+    float xx = 0;
+    for ( size_t i = xStart; i < xStart+frameSize; ++i) {
+        xx += (pitchBuf[i] * pitchBuf[i]);
+    }
+
+    float xy = 0;
+    for ( size_t i = xStart; i < xStart+frameSize; ++i) {
+        xy += (pitchBuf[i] * pitchBuf[i-pitchIdx0]);
+    }
+
+    vec1D32F yyLookup (maxPeriod+1, 0);
+    yyLookup[0] = xx;
+    float yy = xx;
+
+    for ( size_t i = 1; i < yyLookup.size(); ++i) {
+        yy = yy + (pitchBuf[xStart-i] * pitchBuf[xStart-i]) -
+                (pitchBuf[xStart+frameSize-i] * pitchBuf[xStart+frameSize-i]);
+        yyLookup[i] = std::max(0.0f, yy);
+    }
+
+    yy = yyLookup[pitchIdx0];
+    float bestXy = xy;
+    float bestYy = yy;
+
+    float g = this->ComputePitchGain(xy, xx, yy);
+    float g0 = g;
+
+    /* Look for any pitch at pitchIndex/k. */
+    for ( size_t k = 2; k < 16; ++k) {
+        size_t pitchIdx1 = (2*pitchIdx0+k) / (2*k);
+        if (pitchIdx1 < minPeriod) {
+            break;
+        }
+
+        size_t pitchIdx1b;
+        /* Look for another strong correlation at T1b. */
+        if (k == 2) {
+            if ((pitchIdx1 + pitchIdx0) > maxPeriod) {
+                pitchIdx1b = pitchIdx0;
+            } else {
+                pitchIdx1b = pitchIdx0 + pitchIdx1;
+            }
+        } else {
+            pitchIdx1b = (2*(secondCheck[k])*pitchIdx0 + k) / (2*k);
+        }
+
+        xy = 0;
+        for ( size_t i = xStart; i < xStart+frameSize; ++i) {
+            xy += (pitchBuf[i] * pitchBuf[i-pitchIdx1]);
+        }
+
+        float xy2 = 0;
+        for ( size_t i = xStart; i < xStart+frameSize; ++i) {
+            xy2 += (pitchBuf[i] * pitchBuf[i-pitchIdx1b]);
+        }
+        xy = 0.5f * (xy + xy2);
+        VERIFY(pitchIdx1b < maxPeriod+1);
+        yy = 0.5f * (yyLookup[pitchIdx1] + yyLookup[pitchIdx1b]);
+
+        float g1 = this->ComputePitchGain(xy, xx, yy);
+
+        float cont;
+        if (std::abs(pitchIdx1-lastPeriod) <= 1) {
+            cont = lastGain;
+        } else if (std::abs(pitchIdx1-lastPeriod) <= 2 and 5*k*k < pitchIdx0) {
+            cont = 0.5f*lastGain;
+        } else {
+            cont = 0.0f;
+        }
+
+        float thresh = std::max(0.3, 0.7*g0-cont);
+
+        /* Bias against very high pitch (very short period) to avoid false-positives
+         * due to short-term correlation */
+        if (pitchIdx1 < 3*minPeriod) {
+            thresh = std::max(0.4, 0.85*g0-cont);
+        } else if (pitchIdx1 < 2*minPeriod) {
+            thresh = std::max(0.5, 0.9*g0-cont);
+        }
+        if (g1 > thresh) {
+            bestXy = xy;
+            bestYy = yy;
+            pitchIdx = pitchIdx1;
+            g = g1;
+        }
+    }
+
+    bestXy = std::max(0.0f, bestXy);
+    float pg;
+    if (bestYy <= bestXy) {
+        pg = 1.0;
+    } else {
+        pg = bestXy/(bestYy+1);
+    }
+
+    std::array<float, 3> xCorr {0};
+    for ( size_t k = 0; k < 3; ++k ) {
+        for ( size_t i = xStart; i < xStart+frameSize; ++i) {
+            xCorr[k] += (pitchBuf[i] * pitchBuf[i-(pitchIdx+k-1)]);
+        }
+    }
+
+    size_t offset;
+    if ((xCorr[2]-xCorr[0]) > 0.7*(xCorr[1]-xCorr[0])) {
+        offset = 1;
+    } else if ((xCorr[0]-xCorr[2]) > 0.7*(xCorr[1]-xCorr[2])) {
+        offset = -1;
+    } else {
+        offset = 0;
+    }
+
+    if (pg > g) {
+        pg = g;
+    }
+
+    pitchIdx0_ = 2*pitchIdx + offset;
+
+    if (pitchIdx0_ < minPeriod0) {
+        pitchIdx0_ = minPeriod0;
+    }
+
+    this->m_lastPeriod = pitchIdx0_;
+    this->m_lastGain = pg;
+
+    return this->m_lastPeriod;
+}
+
+float RNNoiseFeatureProcessor::ComputePitchGain(float xy, float xx, float yy)
+{
+    return xy / math::MathUtils::SqrtF32(1+xx*yy);
+}
+
+void RNNoiseFeatureProcessor::AutoCorr(
+    const vec1D32F& x,
+    vec1D32F& ac,
+    size_t lag,
+    size_t n)
+{
+    if (n < lag) {
+        printf_err("Invalid parameters for AutoCorr\n");
+        return;
+    }
+
+    auto fastN = n - lag;
+
+    /* Auto-correlation - can be done by PlatformMath functions */
+    this->PitchXCorr(x, x, ac, fastN, lag + 1);
+
+    /* Modify auto-correlation by summing with auto-correlation for different lags. */
+    for (size_t k = 0; k < lag + 1; k++) {
+        float d = 0;
+        for (size_t i = k + fastN; i < n; i++) {
+            d += x[i] * x[i - k];
+        }
+        ac[k] += d;
+    }
+}
+
+
+void RNNoiseFeatureProcessor::PitchXCorr(
+    const vec1D32F& x,
+    const vec1D32F& y,
+    vec1D32F& xCorr,
+    size_t len,
+    size_t maxPitch)
+{
+    for (size_t i = 0; i < maxPitch; i++) {
+        float sum = 0;
+        for (size_t j = 0; j < len; j++) {
+            sum += x[j] * y[i + j];
+        }
+        xCorr[i] = sum;
+    }
+}
+
+/* Linear predictor coefficients */
+void RNNoiseFeatureProcessor::LPC(
+    const vec1D32F& correlation,
+    int32_t p,
+    vec1D32F& lpc)
+{
+    auto error = correlation[0];
+
+    if (error != 0) {
+        for (int i = 0; i < p; i++) {
+
+            /* Sum up this iteration's reflection coefficient */
+            float rr = 0;
+            for (int j = 0; j < i; j++) {
+                rr += lpc[j] * correlation[i - j];
+            }
+
+            rr += correlation[i + 1];
+            auto r = -rr / error;
+
+            /* Update LP coefficients and total error */
+            lpc[i] = r;
+            for (int j = 0; j < ((i + 1) >> 1); j++) {
+                auto tmp1 = lpc[j];
+                auto tmp2 = lpc[i - 1 - j];
+                lpc[j] = tmp1 + (r * tmp2);
+                lpc[i - 1 - j] = tmp2 + (r * tmp1);
+            }
+
+            error = error - (r * r * error);
+
+            /* Bail out once we get 30dB gain */
+            if (error < (0.001 * correlation[0])) {
+                break;
+            }
+        }
+    }
+}
+
+void RNNoiseFeatureProcessor::Fir5(
+    const vec1D32F &num,
+    uint32_t N,
+    vec1D32F &x)
+{
+    auto num0 = num[0];
+    auto num1 = num[1];
+    auto num2 = num[2];
+    auto num3 = num[3];
+    auto num4 = num[4];
+    auto mem0 = 0;
+    auto mem1 = 0;
+    auto mem2 = 0;
+    auto mem3 = 0;
+    auto mem4 = 0;
+    for (uint32_t i = 0; i < N; i++)
+    {
+        auto sum_ = x[i] +  (num0 * mem0) + (num1 * mem1) +
+                    (num2 * mem2) + (num3 * mem3) + (num4 * mem4);
+        mem4 = mem3;
+        mem3 = mem2;
+        mem2 = mem1;
+        mem1 = mem0;
+        mem0 = x[i];
+        x[i] = sum_;
+    }
+}
+
+void RNNoiseFeatureProcessor::PitchFilter(FrameFeatures &features, vec1D32F &gain) {
+    std::vector<float> r(NB_BANDS, 0);
+    std::vector<float> rf(FREQ_SIZE, 0);
+    std::vector<float> newE(NB_BANDS);
+
+    for (size_t i = 0; i < NB_BANDS; i++) {
+        if (features.m_Exp[i] > gain[i]) {
+            r[i] = 1;
+        } else {
+
+
+            r[i] = std::pow(features.m_Exp[i], 2) * (1 - std::pow(gain[i], 2)) /
+                   (.001 + std::pow(gain[i], 2) * (1 - std::pow(features.m_Exp[i], 2)));
+        }
+
+
+        r[i] = math::MathUtils::SqrtF32(std::min(1.0f, std::max(0.0f, r[i])));
+        r[i] *= math::MathUtils::SqrtF32(features.m_Ex[i] / (1e-8f + features.m_Ep[i]));
+    }
+
+    InterpBandGain(rf, r);
+    for (size_t i = 0; i < FREQ_SIZE - 1; i++) {
+        features.m_fftX[2 * i] += rf[i] * features.m_fftP[2 * i];  /* Real. */
+        features.m_fftX[2 * i + 1] += rf[i] * features.m_fftP[2 * i + 1];  /* Imaginary. */
+
+    }
+    ComputeBandEnergy(features.m_fftX, newE);
+    std::vector<float> norm(NB_BANDS);
+    std::vector<float> normf(FRAME_SIZE, 0);
+    for (size_t i = 0; i < NB_BANDS; i++) {
+        norm[i] = math::MathUtils::SqrtF32(features.m_Ex[i] / (1e-8f + newE[i]));
+    }
+
+    InterpBandGain(normf, norm);
+    for (size_t i = 0; i < FREQ_SIZE - 1; i++) {
+        features.m_fftX[2 * i] *= normf[i];  /* Real. */
+        features.m_fftX[2 * i + 1] *= normf[i];  /* Imaginary. */
+
+    }
+}
+
+void RNNoiseFeatureProcessor::FrameSynthesis(vec1D32F& outFrame, vec1D32F& fftY) {
+    std::vector<float> x(WINDOW_SIZE, 0);
+    InverseTransform(x, fftY);
+    ApplyWindow(x);
+    for (size_t i = 0; i < FRAME_SIZE; i++) {
+        outFrame[i] = x[i] + m_synthesisMem[i];
+    }
+    memcpy((m_synthesisMem.data()), &x[FRAME_SIZE], FRAME_SIZE*sizeof(float));
+}
+
+void RNNoiseFeatureProcessor::InterpBandGain(vec1D32F& g, vec1D32F& bandE) {
+    for (size_t i = 0; i < NB_BANDS - 1; i++) {
+        int bandSize = (m_eband5ms[i + 1] - m_eband5ms[i]) << FRAME_SIZE_SHIFT;
+        for (int j = 0; j < bandSize; j++) {
+            float frac = static_cast<float>(j) / bandSize;
+            g[(m_eband5ms[i] << FRAME_SIZE_SHIFT) + j] = (1 - frac) * bandE[i] + frac * bandE[i + 1];
+        }
+    }
+}
+
+void RNNoiseFeatureProcessor::InverseTransform(vec1D32F& out, vec1D32F& fftXIn) {
+
+    std::vector<float> x(WINDOW_SIZE * 2);  /* This is complex. */
+    vec1D32F newFFT;  /* This is complex. */
+
+    size_t i;
+    for (i = 0; i < FREQ_SIZE * 2; i++) {
+        x[i] = fftXIn[i];
+    }
+    for (i = FREQ_SIZE; i < WINDOW_SIZE; i++) {
+        x[2 * i] = x[2 * (WINDOW_SIZE - i)];  /* Real. */
+        x[2 * i + 1] = -x[2 * (WINDOW_SIZE - i) + 1];  /* Imaginary. */
+    }
+
+    constexpr uint32_t numFFt = 2 * FRAME_SIZE;
+    static_assert(numFFt != 0, "numFFt cannot be 0!");
+
+    vec1D32F fftOut = vec1D32F(x.size(), 0);
+    math::MathUtils::FftF32(x,fftOut, m_fftInstCmplx);
+
+    /* Normalize. */
+    for (auto &f: fftOut) {
+        f /= numFFt;
+    }
+
+    out[0] = WINDOW_SIZE * fftOut[0];  /* Real. */
+    for (i = 1; i < WINDOW_SIZE; i++) {
+        out[i] = WINDOW_SIZE * fftOut[(WINDOW_SIZE * 2) - (2 * i)];  /* Real. */
+    }
+}
+
+
+} /* namespace rnn */
+} /* namespace app */
+} /* namspace arm */
diff --git a/source/application/api/use_case/noise_reduction/src/RNNoiseModel.cc b/source/application/api/use_case/noise_reduction/src/RNNoiseModel.cc
new file mode 100644
index 0000000..457cda9
--- /dev/null
+++ b/source/application/api/use_case/noise_reduction/src/RNNoiseModel.cc
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "RNNoiseModel.hpp"
+#include "log_macros.h"
+
+const tflite::MicroOpResolver& arm::app::RNNoiseModel::GetOpResolver()
+{
+    return this->m_opResolver;
+}
+
+bool arm::app::RNNoiseModel::EnlistOperations()
+{
+    this->m_opResolver.AddUnpack();
+    this->m_opResolver.AddFullyConnected();
+    this->m_opResolver.AddSplit();
+    this->m_opResolver.AddSplitV();
+    this->m_opResolver.AddAdd();
+    this->m_opResolver.AddLogistic();
+    this->m_opResolver.AddMul();
+    this->m_opResolver.AddSub();
+    this->m_opResolver.AddTanh();
+    this->m_opResolver.AddPack();
+    this->m_opResolver.AddReshape();
+    this->m_opResolver.AddQuantize();
+    this->m_opResolver.AddConcatenation();
+    this->m_opResolver.AddRelu();
+
+    if (kTfLiteOk == this->m_opResolver.AddEthosU()) {
+        info("Added %s support to op resolver\n",
+            tflite::GetString_ETHOSU());
+    } else {
+        printf_err("Failed to add Arm NPU support to op resolver.");
+        return false;
+    }
+    return true;
+}
+
+bool arm::app::RNNoiseModel::RunInference()
+{
+    return Model::RunInference();
+}
+
+void arm::app::RNNoiseModel::ResetGruState()
+{
+    for (auto& stateMapping: this->m_gruStateMap) {
+        TfLiteTensor* inputGruStateTensor = this->GetInputTensor(stateMapping.second);
+        auto* inputGruState = tflite::GetTensorData<int8_t>(inputGruStateTensor);
+        /* Initial value of states is 0, but this is affected by quantization zero point. */
+        auto quantParams = arm::app::GetTensorQuantParams(inputGruStateTensor);
+        memset(inputGruState, quantParams.offset, inputGruStateTensor->bytes);
+    }
+}
+
+bool arm::app::RNNoiseModel::CopyGruStates()
+{
+    std::vector<std::pair<size_t, std::vector<int8_t>>> tempOutGruStates;
+    /* Saving output states before copying them to input states to avoid output states modification in the tensor.
+     * tflu shares input and output tensors memory, thus writing to input tensor can change output tensor values. */
+    for (auto& stateMapping: this->m_gruStateMap) {
+        TfLiteTensor* outputGruStateTensor = this->GetOutputTensor(stateMapping.first);
+        std::vector<int8_t> tempOutGruState(outputGruStateTensor->bytes);
+        auto* outGruState = tflite::GetTensorData<int8_t>(outputGruStateTensor);
+        memcpy(tempOutGruState.data(), outGruState, outputGruStateTensor->bytes);
+        /* Index of the input tensor and the data to copy. */
+        tempOutGruStates.emplace_back(stateMapping.second, std::move(tempOutGruState));
+    }
+    /* Updating input GRU states with saved GRU output states. */
+    for (auto& stateMapping: tempOutGruStates) {
+        auto outputGruStateTensorData = stateMapping.second;
+        TfLiteTensor* inputGruStateTensor = this->GetInputTensor(stateMapping.first);
+        if (outputGruStateTensorData.size() != inputGruStateTensor->bytes) {
+            printf_err("Unexpected number of bytes for GRU state mapping. Input = %zuz, output = %zuz.\n",
+                       inputGruStateTensor->bytes,
+                       outputGruStateTensorData.size());
+            return false;
+        }
+        auto* inputGruState = tflite::GetTensorData<int8_t>(inputGruStateTensor);
+        auto* outGruState = outputGruStateTensorData.data();
+        memcpy(inputGruState, outGruState, inputGruStateTensor->bytes);
+    }
+    return true;
+}
diff --git a/source/application/api/use_case/noise_reduction/src/RNNoiseProcessing.cc b/source/application/api/use_case/noise_reduction/src/RNNoiseProcessing.cc
new file mode 100644
index 0000000..f6a3ec4
--- /dev/null
+++ b/source/application/api/use_case/noise_reduction/src/RNNoiseProcessing.cc
@@ -0,0 +1,100 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "RNNoiseProcessing.hpp"
+#include "log_macros.h"
+
+namespace arm {
+namespace app {
+
+    RNNoisePreProcess::RNNoisePreProcess(TfLiteTensor* inputTensor,
+            std::shared_ptr<rnn::RNNoiseFeatureProcessor> featureProcessor, std::shared_ptr<rnn::FrameFeatures> frameFeatures)
+    :   m_inputTensor{inputTensor},
+        m_featureProcessor{featureProcessor},
+        m_frameFeatures{frameFeatures}
+    {}
+
+    bool RNNoisePreProcess::DoPreProcess(const void* data, size_t inputSize)
+    {
+        if (data == nullptr) {
+            printf_err("Data pointer is null");
+            return false;
+        }
+
+        auto input = static_cast<const int16_t*>(data);
+        this->m_audioFrame = rnn::vec1D32F(input, input + inputSize);
+        m_featureProcessor->PreprocessFrame(this->m_audioFrame.data(), inputSize, *this->m_frameFeatures);
+
+        QuantizeAndPopulateInput(this->m_frameFeatures->m_featuresVec,
+                this->m_inputTensor->params.scale, this->m_inputTensor->params.zero_point,
+                this->m_inputTensor);
+
+        debug("Input tensor populated \n");
+
+        return true;
+    }
+
+    void RNNoisePreProcess::QuantizeAndPopulateInput(rnn::vec1D32F& inputFeatures,
+            const float quantScale, const int quantOffset,
+            TfLiteTensor* inputTensor)
+    {
+        const float minVal = std::numeric_limits<int8_t>::min();
+        const float maxVal = std::numeric_limits<int8_t>::max();
+
+        auto* inputTensorData = tflite::GetTensorData<int8_t>(inputTensor);
+
+        for (size_t i=0; i < inputFeatures.size(); ++i) {
+            float quantValue = ((inputFeatures[i] / quantScale) + quantOffset);
+            inputTensorData[i] = static_cast<int8_t>(std::min<float>(std::max<float>(quantValue, minVal), maxVal));
+        }
+    }
+
+    RNNoisePostProcess::RNNoisePostProcess(TfLiteTensor* outputTensor,
+            std::vector<int16_t>& denoisedAudioFrame,
+            std::shared_ptr<rnn::RNNoiseFeatureProcessor> featureProcessor,
+            std::shared_ptr<rnn::FrameFeatures> frameFeatures)
+    :   m_outputTensor{outputTensor},
+        m_denoisedAudioFrame{denoisedAudioFrame},
+        m_featureProcessor{featureProcessor},
+        m_frameFeatures{frameFeatures}
+        {
+            this->m_denoisedAudioFrameFloat.reserve(denoisedAudioFrame.size());
+            this->m_modelOutputFloat.resize(outputTensor->bytes);
+        }
+
+    bool RNNoisePostProcess::DoPostProcess()
+    {
+        const auto* outputData = tflite::GetTensorData<int8_t>(this->m_outputTensor);
+        auto outputQuantParams = GetTensorQuantParams(this->m_outputTensor);
+
+        for (size_t i = 0; i < this->m_outputTensor->bytes; ++i) {
+            this->m_modelOutputFloat[i] = (static_cast<float>(outputData[i]) - outputQuantParams.offset)
+                                  * outputQuantParams.scale;
+        }
+
+        this->m_featureProcessor->PostProcessFrame(this->m_modelOutputFloat,
+                *this->m_frameFeatures, this->m_denoisedAudioFrameFloat);
+
+        for (size_t i = 0; i < this->m_denoisedAudioFrame.size(); ++i) {
+            this->m_denoisedAudioFrame[i] = static_cast<int16_t>(
+                    std::roundf(this->m_denoisedAudioFrameFloat[i]));
+        }
+
+        return true;
+    }
+
+} /* namespace app */
+} /* namespace arm */
\ No newline at end of file
diff --git a/source/application/api/use_case/object_detection/CMakeLists.txt b/source/application/api/use_case/object_detection/CMakeLists.txt
new file mode 100644
index 0000000..797ff55
--- /dev/null
+++ b/source/application/api/use_case/object_detection/CMakeLists.txt
@@ -0,0 +1,40 @@
+#----------------------------------------------------------------------------
+#  Copyright (c) 2022 Arm Limited. All rights reserved.
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#----------------------------------------------------------------------------
+#########################################################
+#             OBJECT DETECTION API library              #
+#########################################################
+cmake_minimum_required(VERSION 3.15.6)
+
+set(OBJECT_DETECTION_API_TARGET object_detection_api)
+project(${OBJECT_DETECTION_API_TARGET}
+        DESCRIPTION     "Object detection use case API library"
+        LANGUAGES       C CXX)
+
+# Create static library
+add_library(${OBJECT_DETECTION_API_TARGET} STATIC
+        src/DetectorPreProcessing.cc
+        src/DetectorPostProcessing.cc
+        src/YoloFastestModel.cc)
+
+target_include_directories(${OBJECT_DETECTION_API_TARGET} PUBLIC include)
+
+target_link_libraries(${OBJECT_DETECTION_API_TARGET} PUBLIC common_api)
+
+message(STATUS "*******************************************************")
+message(STATUS "Library                                : " ${OBJECT_DETECTION_API_TARGET})
+message(STATUS "CMAKE_SYSTEM_PROCESSOR                 : " ${CMAKE_SYSTEM_PROCESSOR})
+message(STATUS "*******************************************************")
diff --git a/source/application/api/use_case/object_detection/include/DetectionResult.hpp b/source/application/api/use_case/object_detection/include/DetectionResult.hpp
new file mode 100644
index 0000000..aa74d90
--- /dev/null
+++ b/source/application/api/use_case/object_detection/include/DetectionResult.hpp
@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef DETECTION_RESULT_HPP
+#define DETECTION_RESULT_HPP
+
+
+namespace arm {
+namespace app {
+namespace object_detection {
+
+    /**
+     * @brief   Class representing a single detection result.
+     */
+    class DetectionResult {
+    public:
+        /**
+         * @brief       Constructor
+         * @param[in]   normalisedVal   Result normalized value
+         * @param[in]   x0              Top corner x starting point
+         * @param[in]   y0              Top corner y starting point
+         * @param[in]   w               Detection result width
+         * @param[in]   h               Detection result height
+         **/
+        DetectionResult(double normalisedVal,int x0,int y0, int w,int h) :
+                m_normalisedVal(normalisedVal),
+                m_x0(x0),
+                m_y0(y0),
+                m_w(w),
+                m_h(h)
+            {
+            }
+
+        DetectionResult() = default;
+        ~DetectionResult() = default;
+
+        double  m_normalisedVal{0.0};
+        int     m_x0{0};
+        int     m_y0{0};
+        int     m_w{0};
+        int     m_h{0};
+    };
+
+} /* namespace object_detection */
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* DETECTION_RESULT_HPP */
diff --git a/source/application/api/use_case/object_detection/include/DetectorPostProcessing.hpp b/source/application/api/use_case/object_detection/include/DetectorPostProcessing.hpp
new file mode 100644
index 0000000..30bc123
--- /dev/null
+++ b/source/application/api/use_case/object_detection/include/DetectorPostProcessing.hpp
@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef DETECTOR_POST_PROCESSING_HPP
+#define DETECTOR_POST_PROCESSING_HPP
+
+#include "ImageUtils.hpp"
+#include "DetectionResult.hpp"
+#include "YoloFastestModel.hpp"
+#include "BaseProcessing.hpp"
+
+#include <forward_list>
+
+namespace arm {
+namespace app {
+
+namespace object_detection {
+
+    struct Branch {
+        int resolution;
+        int numBox;
+        const float* anchor;
+        int8_t* modelOutput;
+        float scale;
+        int zeroPoint;
+        size_t size;
+    };
+
+    struct Network {
+        int inputWidth;
+        int inputHeight;
+        int numClasses;
+        std::vector<Branch> branches;
+        int topN;
+    };
+
+} /* namespace object_detection */
+
+    /**
+     * @brief   Post-processing class for Object Detection use case.
+     *          Implements methods declared by BasePostProcess and anything else needed
+     *          to populate result vector.
+     */
+    class DetectorPostProcess : public BasePostProcess {
+    public:
+        /**
+         * @brief        Constructor.
+         * @param[in]    outputTensor0   Pointer to the TFLite Micro output Tensor at index 0.
+         * @param[in]    outputTensor1   Pointer to the TFLite Micro output Tensor at index 1.
+         * @param[out]   results         Vector of detected results.
+         * @param[in]    inputImgRows    Number of rows in the input image.
+         * @param[in]    inputImgCols    Number of columns in the input image.
+         * @param[in]    threshold       Post-processing threshold.
+         * @param[in]    nms             Non-maximum Suppression threshold.
+         * @param[in]    numClasses      Number of classes.
+         * @param[in]    topN            Top N for each class.
+         **/
+        explicit DetectorPostProcess(TfLiteTensor* outputTensor0,
+                                     TfLiteTensor* outputTensor1,
+                                     std::vector<object_detection::DetectionResult>& results,
+                                     int inputImgRows,
+                                     int inputImgCols,
+                                     float threshold = 0.5f,
+                                     float nms = 0.45f,
+                                     int numClasses = 1,
+                                     int topN = 0);
+
+        /**
+         * @brief    Should perform YOLO post-processing of the result of inference then
+         *           populate Detection result data for any later use.
+         * @return   true if successful, false otherwise.
+         **/
+        bool DoPostProcess() override;
+
+    private:
+        TfLiteTensor* m_outputTensor0;     /* Output tensor index 0 */
+        TfLiteTensor* m_outputTensor1;     /* Output tensor index 1 */
+        std::vector<object_detection::DetectionResult>& m_results;  /* Single inference results. */
+        int m_inputImgRows;                /* Number of rows for model input. */
+        int m_inputImgCols;                /* Number of cols for model input. */
+        float m_threshold;                 /* Post-processing threshold. */
+        float m_nms;                       /* NMS threshold. */
+        int   m_numClasses;                /* Number of classes. */
+        int   m_topN;                      /* TopN. */
+        object_detection::Network m_net;   /* YOLO network object. */
+
+        /**
+         * @brief       Insert the given Detection in the list.
+         * @param[in]   detections   List of detections.
+         * @param[in]   det          Detection to be inserted.
+         **/
+        void InsertTopNDetections(std::forward_list<image::Detection>& detections, image::Detection& det);
+
+        /**
+         * @brief        Given a Network calculate the detection boxes.
+         * @param[in]    net           Network.
+         * @param[in]    imageWidth    Original image width.
+         * @param[in]    imageHeight   Original image height.
+         * @param[in]    threshold     Detections threshold.
+         * @param[out]   detections    Detection boxes.
+         **/
+        void GetNetworkBoxes(object_detection::Network& net,
+                             int imageWidth,
+                             int imageHeight,
+                             float threshold,
+                             std::forward_list<image::Detection>& detections);
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* DETECTOR_POST_PROCESSING_HPP */
diff --git a/source/application/api/use_case/object_detection/include/DetectorPreProcessing.hpp b/source/application/api/use_case/object_detection/include/DetectorPreProcessing.hpp
new file mode 100644
index 0000000..4936048
--- /dev/null
+++ b/source/application/api/use_case/object_detection/include/DetectorPreProcessing.hpp
@@ -0,0 +1,60 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef DETECTOR_PRE_PROCESSING_HPP
+#define DETECTOR_PRE_PROCESSING_HPP
+
+#include "BaseProcessing.hpp"
+#include "Classifier.hpp"
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   Pre-processing class for Object detection use case.
+     *          Implements methods declared by BasePreProcess and anything else needed
+     *          to populate input tensors ready for inference.
+     */
+    class DetectorPreProcess : public BasePreProcess {
+
+    public:
+        /**
+         * @brief       Constructor
+         * @param[in]   inputTensor     Pointer to the TFLite Micro input Tensor.
+         * @param[in]   rgb2Gray        Convert image from 3 channel RGB to 1 channel grayscale.
+         * @param[in]   convertToInt8   Convert the image from uint8 to int8 range.
+         **/
+        explicit DetectorPreProcess(TfLiteTensor* inputTensor, bool rgb2Gray, bool convertToInt8);
+
+        /**
+         * @brief       Should perform pre-processing of 'raw' input image data and load it into
+         *              TFLite Micro input tensor ready for inference
+         * @param[in]   input      Pointer to the data that pre-processing will work on.
+         * @param[in]   inputSize  Size of the input data.
+         * @return      true if successful, false otherwise.
+         **/
+        bool DoPreProcess(const void* input, size_t inputSize) override;
+
+    private:
+        TfLiteTensor* m_inputTensor;
+        bool m_rgb2Gray;
+        bool m_convertToInt8;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* DETECTOR_PRE_PROCESSING_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/object_detection/include/YoloFastestModel.hpp b/source/application/api/use_case/object_detection/include/YoloFastestModel.hpp
new file mode 100644
index 0000000..4c64433
--- /dev/null
+++ b/source/application/api/use_case/object_detection/include/YoloFastestModel.hpp
@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef YOLO_FASTEST_MODEL_HPP
+#define YOLO_FASTEST_MODEL_HPP
+
+#include "Model.hpp"
+
+extern const int originalImageSize;
+extern const int channelsImageDisplayed;
+extern const float anchor1[];
+extern const float anchor2[];
+
+namespace arm {
+namespace app {
+
+    class YoloFastestModel : public Model {
+
+    public:
+        /* Indices for the expected model - based on input tensor shape */
+        static constexpr uint32_t ms_inputRowsIdx     = 1;
+        static constexpr uint32_t ms_inputColsIdx     = 2;
+        static constexpr uint32_t ms_inputChannelsIdx = 3;
+
+    protected:
+        /** @brief   Gets the reference to op resolver interface class. */
+        const tflite::MicroOpResolver& GetOpResolver() override;
+
+        /** @brief   Adds operations to the op resolver instance. */
+        bool EnlistOperations() override;
+
+    private:
+        /* Maximum number of individual operations that can be enlisted. */
+        static constexpr int ms_maxOpCnt = 8;
+
+        /* A mutable op resolver instance. */
+        tflite::MicroMutableOpResolver<ms_maxOpCnt> m_opResolver;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* YOLO_FASTEST_MODEL_HPP */
diff --git a/source/application/api/use_case/object_detection/src/DetectorPostProcessing.cc b/source/application/api/use_case/object_detection/src/DetectorPostProcessing.cc
new file mode 100644
index 0000000..fb1606a
--- /dev/null
+++ b/source/application/api/use_case/object_detection/src/DetectorPostProcessing.cc
@@ -0,0 +1,240 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "DetectorPostProcessing.hpp"
+#include "PlatformMath.hpp"
+
+#include <cmath>
+
+namespace arm {
+namespace app {
+
+    DetectorPostProcess::DetectorPostProcess(
+        TfLiteTensor* modelOutput0,
+        TfLiteTensor* modelOutput1,
+        std::vector<object_detection::DetectionResult>& results,
+        int inputImgRows,
+        int inputImgCols,
+        const float threshold,
+        const float nms,
+        int numClasses,
+        int topN)
+        :   m_outputTensor0{modelOutput0},
+            m_outputTensor1{modelOutput1},
+            m_results{results},
+            m_inputImgRows{inputImgRows},
+            m_inputImgCols{inputImgCols},
+            m_threshold(threshold),
+            m_nms(nms),
+            m_numClasses(numClasses),
+            m_topN(topN)
+{
+    /* Init PostProcessing */
+    this->m_net =
+    object_detection::Network {
+        .inputWidth = inputImgCols,
+        .inputHeight = inputImgRows,
+        .numClasses = numClasses,
+        .branches = {
+            object_detection::Branch {
+                        .resolution = inputImgCols/32,
+                        .numBox = 3,
+                        .anchor = anchor1,
+                        .modelOutput = this->m_outputTensor0->data.int8,
+                        .scale = (static_cast<TfLiteAffineQuantization*>(
+                                this->m_outputTensor0->quantization.params))->scale->data[0],
+                        .zeroPoint = (static_cast<TfLiteAffineQuantization*>(
+                                this->m_outputTensor0->quantization.params))->zero_point->data[0],
+                        .size = this->m_outputTensor0->bytes
+            },
+            object_detection::Branch {
+                    .resolution = inputImgCols/16,
+                    .numBox = 3,
+                    .anchor = anchor2,
+                    .modelOutput = this->m_outputTensor1->data.int8,
+                    .scale = (static_cast<TfLiteAffineQuantization*>(
+                            this->m_outputTensor1->quantization.params))->scale->data[0],
+                    .zeroPoint = (static_cast<TfLiteAffineQuantization*>(
+                            this->m_outputTensor1->quantization.params))->zero_point->data[0],
+                    .size = this->m_outputTensor1->bytes
+            }
+        },
+        .topN = m_topN
+    };
+    /* End init */
+}
+
+bool DetectorPostProcess::DoPostProcess()
+{
+    /* Start postprocessing */
+    int originalImageWidth = originalImageSize;
+    int originalImageHeight = originalImageSize;
+
+    std::forward_list<image::Detection> detections;
+    GetNetworkBoxes(this->m_net, originalImageWidth, originalImageHeight, m_threshold, detections);
+
+    /* Do nms */
+    CalculateNMS(detections, this->m_net.numClasses, m_nms);
+
+    for (auto& it: detections) {
+        float xMin = it.bbox.x - it.bbox.w / 2.0f;
+        float xMax = it.bbox.x + it.bbox.w / 2.0f;
+        float yMin = it.bbox.y - it.bbox.h / 2.0f;
+        float yMax = it.bbox.y + it.bbox.h / 2.0f;
+
+        if (xMin < 0) {
+            xMin = 0;
+        }
+        if (yMin < 0) {
+            yMin = 0;
+        }
+        if (xMax > originalImageWidth) {
+            xMax = originalImageWidth;
+        }
+        if (yMax > originalImageHeight) {
+            yMax = originalImageHeight;
+        }
+
+        float boxX = xMin;
+        float boxY = yMin;
+        float boxWidth = xMax - xMin;
+        float boxHeight = yMax - yMin;
+
+        for (int j = 0; j < this->m_net.numClasses; ++j) {
+            if (it.prob[j] > 0) {
+
+                object_detection::DetectionResult tmpResult = {};
+                tmpResult.m_normalisedVal = it.prob[j];
+                tmpResult.m_x0 = boxX;
+                tmpResult.m_y0 = boxY;
+                tmpResult.m_w = boxWidth;
+                tmpResult.m_h = boxHeight;
+
+                this->m_results.push_back(tmpResult);
+            }
+        }
+    }
+    return true;
+}
+
+void DetectorPostProcess::InsertTopNDetections(std::forward_list<image::Detection>& detections, image::Detection& det)
+{
+    std::forward_list<image::Detection>::iterator it;
+    std::forward_list<image::Detection>::iterator last_it;
+    for ( it = detections.begin(); it != detections.end(); ++it ) {
+        if(it->objectness > det.objectness)
+            break;
+        last_it = it;
+    }
+    if(it != detections.begin()) {
+        detections.emplace_after(last_it, det);
+        detections.pop_front();
+    }
+}
+
+void DetectorPostProcess::GetNetworkBoxes(
+        object_detection::Network& net,
+        int imageWidth,
+        int imageHeight,
+        float threshold,
+        std::forward_list<image::Detection>& detections)
+{
+    int numClasses = net.numClasses;
+    int num = 0;
+    auto det_objectness_comparator = [](image::Detection& pa, image::Detection& pb) {
+        return pa.objectness < pb.objectness;
+    };
+    for (size_t i = 0; i < net.branches.size(); ++i) {
+        int height   = net.branches[i].resolution;
+        int width    = net.branches[i].resolution;
+        int channel  = net.branches[i].numBox*(5+numClasses);
+
+        for (int h = 0; h < net.branches[i].resolution; h++) {
+            for (int w = 0; w < net.branches[i].resolution; w++) {
+                for (int anc = 0; anc < net.branches[i].numBox; anc++) {
+
+                    /* Objectness score */
+                    int bbox_obj_offset = h * width * channel + w * channel + anc * (numClasses + 5) + 4;
+                    float objectness = math::MathUtils::SigmoidF32(
+                            (static_cast<float>(net.branches[i].modelOutput[bbox_obj_offset])
+                            - net.branches[i].zeroPoint
+                            ) * net.branches[i].scale);
+
+                    if(objectness > threshold) {
+                        image::Detection det;
+                        det.objectness = objectness;
+                        /* Get bbox prediction data for each anchor, each feature point */
+                        int bbox_x_offset = bbox_obj_offset -4;
+                        int bbox_y_offset = bbox_x_offset + 1;
+                        int bbox_w_offset = bbox_x_offset + 2;
+                        int bbox_h_offset = bbox_x_offset + 3;
+                        int bbox_scores_offset = bbox_x_offset + 5;
+
+                        det.bbox.x = (static_cast<float>(net.branches[i].modelOutput[bbox_x_offset])
+                                - net.branches[i].zeroPoint) * net.branches[i].scale;
+                        det.bbox.y = (static_cast<float>(net.branches[i].modelOutput[bbox_y_offset])
+                                - net.branches[i].zeroPoint) * net.branches[i].scale;
+                        det.bbox.w = (static_cast<float>(net.branches[i].modelOutput[bbox_w_offset])
+                                - net.branches[i].zeroPoint) * net.branches[i].scale;
+                        det.bbox.h = (static_cast<float>(net.branches[i].modelOutput[bbox_h_offset])
+                                - net.branches[i].zeroPoint) * net.branches[i].scale;
+
+                        float bbox_x, bbox_y;
+
+                        /* Eliminate grid sensitivity trick involved in YOLOv4 */
+                        bbox_x = math::MathUtils::SigmoidF32(det.bbox.x);
+                        bbox_y = math::MathUtils::SigmoidF32(det.bbox.y);
+                        det.bbox.x = (bbox_x + w) / width;
+                        det.bbox.y = (bbox_y + h) / height;
+
+                        det.bbox.w = std::exp(det.bbox.w) * net.branches[i].anchor[anc*2] / net.inputWidth;
+                        det.bbox.h = std::exp(det.bbox.h) * net.branches[i].anchor[anc*2+1] / net.inputHeight;
+
+                        for (int s = 0; s < numClasses; s++) {
+                            float sig = math::MathUtils::SigmoidF32(
+                                    (static_cast<float>(net.branches[i].modelOutput[bbox_scores_offset + s]) -
+                                    net.branches[i].zeroPoint) * net.branches[i].scale
+                                    ) * objectness;
+                            det.prob.emplace_back((sig > threshold) ? sig : 0);
+                        }
+
+                        /* Correct_YOLO_boxes */
+                        det.bbox.x *= imageWidth;
+                        det.bbox.w *= imageWidth;
+                        det.bbox.y *= imageHeight;
+                        det.bbox.h *= imageHeight;
+
+                        if (num < net.topN || net.topN <=0) {
+                            detections.emplace_front(det);
+                            num += 1;
+                        } else if (num == net.topN) {
+                            detections.sort(det_objectness_comparator);
+                            InsertTopNDetections(detections,det);
+                            num += 1;
+                        } else {
+                            InsertTopNDetections(detections,det);
+                        }
+                    }
+                }
+            }
+        }
+    }
+    if(num > net.topN)
+        num -=1;
+}
+
+} /* namespace app */
+} /* namespace arm */
diff --git a/source/application/api/use_case/object_detection/src/DetectorPreProcessing.cc b/source/application/api/use_case/object_detection/src/DetectorPreProcessing.cc
new file mode 100644
index 0000000..7212046
--- /dev/null
+++ b/source/application/api/use_case/object_detection/src/DetectorPreProcessing.cc
@@ -0,0 +1,52 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "DetectorPreProcessing.hpp"
+#include "ImageUtils.hpp"
+#include "log_macros.h"
+
+namespace arm {
+namespace app {
+
+    DetectorPreProcess::DetectorPreProcess(TfLiteTensor* inputTensor, bool rgb2Gray, bool convertToInt8)
+    :   m_inputTensor{inputTensor},
+        m_rgb2Gray{rgb2Gray},
+        m_convertToInt8{convertToInt8}
+    {}
+
+    bool DetectorPreProcess::DoPreProcess(const void* data, size_t inputSize) {
+        if (data == nullptr) {
+            printf_err("Data pointer is null");
+        }
+
+        auto input = static_cast<const uint8_t*>(data);
+
+        if (this->m_rgb2Gray) {
+            image::RgbToGrayscale(input, this->m_inputTensor->data.uint8, this->m_inputTensor->bytes);
+        } else {
+            std::memcpy(this->m_inputTensor->data.data, input, inputSize);
+        }
+        debug("Input tensor populated \n");
+
+        if (this->m_convertToInt8) {
+            image::ConvertImgToInt8(this->m_inputTensor->data.data, this->m_inputTensor->bytes);
+        }
+
+        return true;
+    }
+
+} /* namespace app */
+} /* namespace arm */
\ No newline at end of file
diff --git a/source/application/api/use_case/object_detection/src/YoloFastestModel.cc b/source/application/api/use_case/object_detection/src/YoloFastestModel.cc
new file mode 100644
index 0000000..e293181
--- /dev/null
+++ b/source/application/api/use_case/object_detection/src/YoloFastestModel.cc
@@ -0,0 +1,45 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "YoloFastestModel.hpp"
+
+#include "log_macros.h"
+
+const tflite::MicroOpResolver& arm::app::YoloFastestModel::GetOpResolver()
+{
+    return this->m_opResolver;
+}
+
+bool arm::app::YoloFastestModel::EnlistOperations()
+{
+    this->m_opResolver.AddDepthwiseConv2D();
+    this->m_opResolver.AddConv2D();
+    this->m_opResolver.AddAdd();
+    this->m_opResolver.AddResizeNearestNeighbor();
+    /*These are needed for UT to work, not needed on FVP */
+    this->m_opResolver.AddPad();
+    this->m_opResolver.AddMaxPool2D();
+    this->m_opResolver.AddConcatenation();
+
+    if (kTfLiteOk == this->m_opResolver.AddEthosU()) {
+        info("Added %s support to op resolver\n",
+            tflite::GetString_ETHOSU());
+    } else {
+        printf_err("Failed to add Arm NPU support to op resolver.");
+        return false;
+    }
+    return true;
+}
diff --git a/source/application/api/use_case/vww/CMakeLists.txt b/source/application/api/use_case/vww/CMakeLists.txt
new file mode 100644
index 0000000..b933d32
--- /dev/null
+++ b/source/application/api/use_case/vww/CMakeLists.txt
@@ -0,0 +1,39 @@
+#----------------------------------------------------------------------------
+#  Copyright (c) 2022 Arm Limited. All rights reserved.
+#  SPDX-License-Identifier: Apache-2.0
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#----------------------------------------------------------------------------
+#########################################################
+#             VISUAL WAKE WORD API library              #
+#########################################################
+cmake_minimum_required(VERSION 3.15.6)
+
+set(VWW_API_TARGET vww_api)
+project(${VWW_API_TARGET}
+        DESCRIPTION     "Visual wake word use case API library"
+        LANGUAGES       C CXX)
+
+# Create static library
+add_library(${VWW_API_TARGET} STATIC
+        src/VisualWakeWordProcessing.cc
+        src/VisualWakeWordModel.cc)
+
+target_include_directories(${VWW_API_TARGET} PUBLIC include)
+
+target_link_libraries(${VWW_API_TARGET} PUBLIC common_api)
+
+message(STATUS "*******************************************************")
+message(STATUS "Library                                : " ${VWW_API_TARGET})
+message(STATUS "CMAKE_SYSTEM_PROCESSOR                 : " ${CMAKE_SYSTEM_PROCESSOR})
+message(STATUS "*******************************************************")
diff --git a/source/application/api/use_case/vww/include/VisualWakeWordModel.hpp b/source/application/api/use_case/vww/include/VisualWakeWordModel.hpp
new file mode 100644
index 0000000..a34b904
--- /dev/null
+++ b/source/application/api/use_case/vww/include/VisualWakeWordModel.hpp
@@ -0,0 +1,50 @@
+/*
+ * Copyright (c) 2021 - 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef VISUAL_WAKE_WORD_MODEL_HPP
+#define VISUAL_WAKE_WORD_MODEL_HPP
+
+#include "Model.hpp"
+
+namespace arm {
+namespace app {
+
+    class VisualWakeWordModel : public Model {
+
+    public:
+        /* Indices for the expected model - based on input tensor shape */
+        static constexpr uint32_t ms_inputRowsIdx     = 1;
+        static constexpr uint32_t ms_inputColsIdx     = 2;
+        static constexpr uint32_t ms_inputChannelsIdx = 3;
+
+    protected:
+        /** @brief   Gets the reference to op resolver interface class. */
+        const tflite::MicroOpResolver& GetOpResolver() override;
+
+        /** @brief   Adds operations to the op resolver instance. */
+        bool EnlistOperations() override;
+    private:
+        /* Maximum number of individual operations that can be enlisted. */
+        static constexpr int ms_maxOpCnt = 7;
+
+        /* A mutable op resolver instance. */
+        tflite::MicroMutableOpResolver<ms_maxOpCnt> m_opResolver;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* VISUAL_WAKE_WORD_MODEL_HPP */
diff --git a/source/application/api/use_case/vww/include/VisualWakeWordProcessing.hpp b/source/application/api/use_case/vww/include/VisualWakeWordProcessing.hpp
new file mode 100644
index 0000000..f9f9d72
--- /dev/null
+++ b/source/application/api/use_case/vww/include/VisualWakeWordProcessing.hpp
@@ -0,0 +1,93 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef VWW_PROCESSING_HPP
+#define VWW_PROCESSING_HPP
+
+#include "BaseProcessing.hpp"
+#include "Model.hpp"
+#include "Classifier.hpp"
+
+namespace arm {
+namespace app {
+
+    /**
+     * @brief   Pre-processing class for Visual Wake Word use case.
+     *          Implements methods declared by BasePreProcess and anything else needed
+     *          to populate input tensors ready for inference.
+     */
+    class VisualWakeWordPreProcess : public BasePreProcess {
+
+    public:
+        /**
+         * @brief       Constructor
+         * @param[in]   inputTensor   Pointer to the TFLite Micro input Tensor.
+         * @param[in]   rgb2Gray      Convert image from 3 channel RGB to 1 channel grayscale.
+         **/
+        explicit VisualWakeWordPreProcess(TfLiteTensor* inputTensor, bool rgb2Gray=true);
+
+        /**
+         * @brief       Should perform pre-processing of 'raw' input image data and load it into
+         *              TFLite Micro input tensors ready for inference
+         * @param[in]   input      Pointer to the data that pre-processing will work on.
+         * @param[in]   inputSize  Size of the input data.
+         * @return      true if successful, false otherwise.
+         **/
+        bool DoPreProcess(const void* input, size_t inputSize) override;
+
+    private:
+        TfLiteTensor* m_inputTensor;
+        bool m_rgb2Gray;
+    };
+
+    /**
+     * @brief   Post-processing class for Visual Wake Word use case.
+     *          Implements methods declared by BasePostProcess and anything else needed
+     *          to populate result vector.
+     */
+    class VisualWakeWordPostProcess : public BasePostProcess {
+
+    private:
+        TfLiteTensor* m_outputTensor;
+        Classifier& m_vwwClassifier;
+        const std::vector<std::string>& m_labels;
+        std::vector<ClassificationResult>& m_results;
+
+    public:
+        /**
+         * @brief       Constructor
+         * @param[in]   outputTensor   Pointer to the TFLite Micro output Tensor.
+         * @param[in]   classifier     Classifier object used to get top N results from classification.
+         * @param[in]   model          Pointer to the VWW classification Model object.
+         * @param[in]   labels         Vector of string labels to identify each output of the model.
+         * @param[out]  results        Vector of classification results to store decoded outputs.
+         **/
+        VisualWakeWordPostProcess(TfLiteTensor* outputTensor, Classifier& classifier,
+                const std::vector<std::string>& labels,
+                std::vector<ClassificationResult>& results);
+
+        /**
+         * @brief    Should perform post-processing of the result of inference then
+         *           populate classification result data for any later use.
+         * @return   true if successful, false otherwise.
+         **/
+        bool DoPostProcess() override;
+    };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* VWW_PROCESSING_HPP */
\ No newline at end of file
diff --git a/source/application/api/use_case/vww/src/VisualWakeWordModel.cc b/source/application/api/use_case/vww/src/VisualWakeWordModel.cc
new file mode 100644
index 0000000..2d8a125
--- /dev/null
+++ b/source/application/api/use_case/vww/src/VisualWakeWordModel.cc
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "VisualWakeWordModel.hpp"
+#include "log_macros.h"
+
+const tflite::MicroOpResolver& arm::app::VisualWakeWordModel::GetOpResolver()
+{
+    return this->m_opResolver;
+}
+
+bool arm::app::VisualWakeWordModel::EnlistOperations()
+{
+    this->m_opResolver.AddDepthwiseConv2D();
+    this->m_opResolver.AddConv2D();
+    this->m_opResolver.AddAveragePool2D();
+    this->m_opResolver.AddReshape();
+    this->m_opResolver.AddPad();
+    this->m_opResolver.AddAdd();
+
+    if (kTfLiteOk == this->m_opResolver.AddEthosU()) {
+        info("Added %s support to op resolver\n",
+            tflite::GetString_ETHOSU());
+    } else {
+        printf_err("Failed to add Arm NPU support to op resolver.");
+        return false;
+    }
+    return true;
+}
diff --git a/source/application/api/use_case/vww/src/VisualWakeWordProcessing.cc b/source/application/api/use_case/vww/src/VisualWakeWordProcessing.cc
new file mode 100644
index 0000000..4ae8a54
--- /dev/null
+++ b/source/application/api/use_case/vww/src/VisualWakeWordProcessing.cc
@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "VisualWakeWordProcessing.hpp"
+
+#include "ImageUtils.hpp"
+#include "VisualWakeWordModel.hpp"
+#include "log_macros.h"
+
+namespace arm {
+namespace app {
+
+    VisualWakeWordPreProcess::VisualWakeWordPreProcess(TfLiteTensor* inputTensor, bool rgb2Gray)
+    :m_inputTensor{inputTensor},
+     m_rgb2Gray{rgb2Gray}
+    {}
+
+    bool VisualWakeWordPreProcess::DoPreProcess(const void* data, size_t inputSize)
+    {
+        if (data == nullptr) {
+            printf_err("Data pointer is null");
+        }
+
+        auto input = static_cast<const uint8_t*>(data);
+
+        uint8_t* unsignedDstPtr = this->m_inputTensor->data.uint8;
+
+        if (this->m_rgb2Gray) {
+            image::RgbToGrayscale(input, unsignedDstPtr, inputSize);
+        } else {
+            std::memcpy(unsignedDstPtr, input, inputSize);
+        }
+
+        /* VWW model pre-processing is image conversion from uint8 to [0,1] float values,
+         * then quantize them with input quantization info. */
+        QuantParams inQuantParams = GetTensorQuantParams(this->m_inputTensor);
+
+        int8_t* signedDstPtr = this->m_inputTensor->data.int8;
+        for (size_t i = 0; i < this->m_inputTensor->bytes; i++) {
+            auto i_data_int8 = static_cast<int8_t>(
+                    ((static_cast<float>(unsignedDstPtr[i]) / 255.0f) / inQuantParams.scale) + inQuantParams.offset
+                    );
+            signedDstPtr[i] = std::min<int8_t>(INT8_MAX, std::max<int8_t>(i_data_int8, INT8_MIN));
+        }
+
+        debug("Input tensor populated \n");
+
+        return true;
+    }
+
+    VisualWakeWordPostProcess::VisualWakeWordPostProcess(TfLiteTensor* outputTensor, Classifier& classifier,
+            const std::vector<std::string>& labels, std::vector<ClassificationResult>& results)
+            :m_outputTensor{outputTensor},
+             m_vwwClassifier{classifier},
+             m_labels{labels},
+             m_results{results}
+    {}
+
+    bool VisualWakeWordPostProcess::DoPostProcess()
+    {
+        return this->m_vwwClassifier.GetClassificationResults(
+                this->m_outputTensor, this->m_results,
+                this->m_labels, 1, true);
+    }
+
+} /* namespace app */
+} /* namespace arm */
\ No newline at end of file