MLECO-3077: Add ASR use case API
* Minor adjustments to doc strings in KWS
* Remove unused score threshold in KWS
Signed-off-by: Richard Burton <richard.burton@arm.com>
Change-Id: Ie1c5bf6f7bdbebb853b6a10cb7ba1c4a1d9a76c9
diff --git a/source/use_case/asr/include/AsrResult.hpp b/source/use_case/asr/include/AsrResult.hpp
index b12ed7d..ed826d0 100644
--- a/source/use_case/asr/include/AsrResult.hpp
+++ b/source/use_case/asr/include/AsrResult.hpp
@@ -25,7 +25,7 @@
namespace app {
namespace asr {
- using ResultVec = std::vector < arm::app::ClassificationResult >;
+ using ResultVec = std::vector<arm::app::ClassificationResult>;
/* Structure for holding ASR result. */
class AsrResult {
diff --git a/source/use_case/asr/include/Wav2LetterModel.hpp b/source/use_case/asr/include/Wav2LetterModel.hpp
index 55395b9..895df2b 100644
--- a/source/use_case/asr/include/Wav2LetterModel.hpp
+++ b/source/use_case/asr/include/Wav2LetterModel.hpp
@@ -36,6 +36,9 @@
static constexpr uint32_t ms_outputRowsIdx = 2;
static constexpr uint32_t ms_outputColsIdx = 3;
+ 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;
diff --git a/source/use_case/asr/include/Wav2LetterPostprocess.hpp b/source/use_case/asr/include/Wav2LetterPostprocess.hpp
index 29eb548..45defa5 100644
--- a/source/use_case/asr/include/Wav2LetterPostprocess.hpp
+++ b/source/use_case/asr/include/Wav2LetterPostprocess.hpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * 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");
@@ -17,93 +17,90 @@
#ifndef ASR_WAV2LETTER_POSTPROCESS_HPP
#define ASR_WAV2LETTER_POSTPROCESS_HPP
-#include "TensorFlowLiteMicro.hpp" /* TensorFlow headers. */
+#include "TensorFlowLiteMicro.hpp" /* TensorFlow headers. */
+#include "BaseProcessing.hpp"
+#include "AsrClassifier.hpp"
+#include "AsrResult.hpp"
#include "log_macros.h"
namespace arm {
namespace app {
-namespace audio {
-namespace asr {
/**
* @brief Helper class to manage tensor post-processing for "wav2letter"
* output.
*/
- class Postprocess {
+ 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] contextLen Left and right context length for
- * output tensor.
- * @param[in] innerLen This is the length of the section
- * between left and right context.
- * @param[in] blankTokenIdx Blank token index.
+ * @brief Constructor
+ * @param[in] outputTensor Pointer to the output Tensor.
+ * @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.
**/
- Postprocess(uint32_t contextLen,
- uint32_t innerLen,
- uint32_t blankTokenIdx);
-
- Postprocess() = delete;
- ~Postprocess() = default;
+ ASRPostProcess(AsrClassifier& classifier, TfLiteTensor* outputTensor,
+ const std::vector<std::string>& labels, asr::ResultVec& result,
+ uint32_t outputContextLen,
+ uint32_t blankTokenIdx, uint32_t reductionAxis);
/**
- * @brief Erases the required part of the tensor based
- * on context lengths set up during initialisation.
- * @param[in] tensor Pointer to the tensor.
- * @param[in] axisIdx Index of the axis on which erase is
- * performed.
- * @param[in] lastIteration Flag to signal this is the
- * last iteration in which case
- * the right context is preserved.
- * @return true if successful, false otherwise.
- */
- bool Invoke(TfLiteTensor* tensor,
- uint32_t axisIdx,
- bool lastIteration = false);
+ * @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:
- uint32_t m_contextLen; /* lengths of left and right contexts. */
- uint32_t m_innerLen; /* Length of 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. */
- /**
- * @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,
- const uint32_t axisIdx) const;
+ 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 Gets the tensor data element size in bytes based
- * on the tensor type.
- * @return Size in bytes, 0 if not supported.
+ * @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.
+ * @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,
- const uint32_t strideSzBytes,
- const bool lastIteration);
-
- /**
- * @brief Erases sections from the data assuming col-wise
- * arrangement along the context axis.
- * @return true if successful, false otherwise.
- */
- static bool EraseSectionsColWise(const uint8_t* ptrData,
- const uint32_t strideSzBytes,
- const bool lastIteration);
+ uint32_t strideSzBytes,
+ bool lastIteration);
};
-} /* namespace asr */
-} /* namespace audio */
} /* namespace app */
} /* namespace arm */
diff --git a/source/use_case/asr/include/Wav2LetterPreprocess.hpp b/source/use_case/asr/include/Wav2LetterPreprocess.hpp
index 13d1589..8c12b3d 100644
--- a/source/use_case/asr/include/Wav2LetterPreprocess.hpp
+++ b/source/use_case/asr/include/Wav2LetterPreprocess.hpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * 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");
@@ -21,49 +21,44 @@
#include "Wav2LetterMfcc.hpp"
#include "AudioUtils.hpp"
#include "DataStructures.hpp"
+#include "BaseProcessing.hpp"
#include "log_macros.h"
namespace arm {
namespace app {
-namespace audio {
-namespace asr {
/* Class to facilitate pre-processing calculation for Wav2Letter model
* for ASR. */
- using AudioWindow = SlidingWindow <const int16_t>;
+ using AudioWindow = audio::SlidingWindow<const int16_t>;
- class Preprocess {
+ class ASRPreProcess : public BasePreProcess {
public:
/**
* @brief Constructor.
- * @param[in] numMfccFeatures Number of MFCC features per window.
- * @param[in] windowLen Number of elements in a window.
- * @param[in] windowStride Stride (in number of elements) for
- * moving the window.
- * @param[in] numMfccVectors Number of MFCC vectors per window.
+ * @param[in] inputTensor Pointer to the TFLite Micro input Tensor.
+ * @param[in] numMfccFeatures Number of MFCC features per window.
+ * @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.
+ * @param[in] mfccWindowStride Number of MFCC vectors that need to be calculated
+ * for an inference.
*/
- Preprocess(
- uint32_t numMfccFeatures,
- uint32_t windowLen,
- uint32_t windowStride,
- uint32_t numMfccVectors);
- Preprocess() = delete;
- ~Preprocess() = default;
+ ASRPreProcess(TfLiteTensor* inputTensor,
+ uint32_t numMfccFeatures,
+ uint32_t audioWindowLen,
+ 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 feature from a given window placed along
+ * 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.
- * @param[in] tensor Tensor to be populated.
* @return true if successful, false in case of error.
*/
- bool Invoke(const int16_t * audioData,
- uint32_t audioDataLen,
- TfLiteTensor * tensor);
+ bool DoPreProcess(const void* audioData, size_t audioDataLen) override;
protected:
/**
@@ -80,32 +75,16 @@
Array2d<float>& delta2);
/**
- * @brief Given a 2D vector of floats, computes the mean.
- * @param[in] vec Vctor of vector of floats.
- * @return Mean value.
- */
- static float GetMean(Array2d<float>& vec);
-
- /**
- * @brief Given a 2D vector of floats, computes the stddev.
- * @param[in] vec Vector of vector of floats.
- * @param[in] mean Mean value of the vector passed in.
- * @return stddev value.
- */
- static float GetStdDev(Array2d<float>& vec,
- const float mean);
-
- /**
- * @brief Given a 2D vector of floats, normalises it using
- * the mean and the stddev.
+ * @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 NormaliseVec(Array2d<float>& vec);
+ static void StandardizeVecF32(Array2d<float>& vec);
/**
- * @brief Normalises the MFCC and delta buffers.
+ * @brief Standardizes all the MFCC and delta buffers to have mean 0 and std. dev 1.
*/
- void Normalise();
+ void Standarize();
/**
* @brief Given the quantisation and data type limits, computes
@@ -139,7 +118,7 @@
*/
template <typename T>
bool Quantise(
- T * outputBuf,
+ T* outputBuf,
const uint32_t outputBufSz,
const float quantScale,
const int quantOffset)
@@ -160,15 +139,15 @@
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_numFeatVectors; ++j) {
+ for (uint32_t j = 0; j < this->m_numFeatureFrames; ++j) {
for (uint32_t i = 0; i < this->m_numMfccFeats; ++i) {
- *outputBufMfcc++ = static_cast<T>(Preprocess::GetQuantElem(
+ *outputBufMfcc++ = static_cast<T>(ASRPreProcess::GetQuantElem(
this->m_mfccBuf(i, j), quantScale,
quantOffset, minVal, maxVal));
- *outputBufD1++ = static_cast<T>(Preprocess::GetQuantElem(
+ *outputBufD1++ = static_cast<T>(ASRPreProcess::GetQuantElem(
this->m_delta1Buf(i, j), quantScale,
quantOffset, minVal, maxVal));
- *outputBufD2++ = static_cast<T>(Preprocess::GetQuantElem(
+ *outputBufD2++ = static_cast<T>(ASRPreProcess::GetQuantElem(
this->m_delta2Buf(i, j), quantScale,
quantOffset, minVal, maxVal));
}
@@ -181,23 +160,22 @@
}
private:
- Wav2LetterMFCC m_mfcc; /* MFCC instance. */
+ 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 */
+ 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_windowLen; /* Window length for MFCC. */
- uint32_t m_windowStride; /* Window stride len for MFCC. */
- uint32_t m_numMfccFeats; /* Number of MFCC features per window. */
- uint32_t m_numFeatVectors; /* Number of m_numMfccFeats. */
- AudioWindow m_window; /* Sliding window. */
+ 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 asr */
-} /* namespace audio */
} /* namespace app */
} /* namespace arm */