MLECO-3173: Add AD, KWS_ASR and Noise reduction use case API's
Signed-off-by: Richard Burton <richard.burton@arm.com>
Change-Id: I36f61ce74bf17f7b327cdae9704a22ca54144f37
diff --git a/source/use_case/kws_asr/include/Wav2LetterPreprocess.hpp b/source/use_case/kws_asr/include/Wav2LetterPreprocess.hpp
index 3609c49..1224c23 100644
--- a/source/use_case/kws_asr/include/Wav2LetterPreprocess.hpp
+++ b/source/use_case/kws_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");
@@ -14,56 +14,51 @@
* See the License for the specific language governing permissions and
* limitations under the License.
*/
-#ifndef KWS_ASR_WAV2LET_PREPROC_HPP
-#define KWS_ASR_WAV2LET_PREPROC_HPP
+#ifndef KWS_ASR_WAV2LETTER_PREPROCESS_HPP
+#define KWS_ASR_WAV2LETTER_PREPROCESS_HPP
#include "Wav2LetterModel.hpp"
#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.
- */
- Preprocess(
- uint32_t numMfccFeatures,
- uint32_t windowLen,
- uint32_t windowStride,
- uint32_t numMfccVectors);
- Preprocess() = delete;
- ~Preprocess() = default;
+ * @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 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:
/**
@@ -73,49 +68,32 @@
* @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.
+ * @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, computes the mean.
- * @param[in] vec Vector 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
* 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.
+ * @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(
@@ -133,44 +111,43 @@
* 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.
+ * @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,
+ T* outputBuf,
const uint32_t outputBufSz,
const float quantScale,
const int quantOffset)
{
- /* Check the output size will for everything. */
+ /* 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;
+ 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();
- /* We need to do a transpose while copying and concatenating
- * the tensor. */
- for (uint32_t j = 0; j < this->m_numFeatVectors; ++j) {
+ /* 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>(this->GetQuantElem(
+ *outputBufMfcc++ = static_cast<T>(AsrPreProcess::GetQuantElem(
this->m_mfccBuf(i, j), quantScale,
quantOffset, minVal, maxVal));
- *outputBufD1++ = static_cast<T>(this->GetQuantElem(
+ *outputBufD1++ = static_cast<T>(AsrPreProcess::GetQuantElem(
this->m_delta1Buf(i, j), quantScale,
quantOffset, minVal, maxVal));
- *outputBufD2++ = static_cast<T>(this->GetQuantElem(
+ *outputBufD2++ = static_cast<T>(AsrPreProcess::GetQuantElem(
this->m_delta2Buf(i, j), quantScale,
quantOffset, minVal, maxVal));
}
@@ -183,24 +160,23 @@
}
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 */
-#endif /* KWS_ASR_WAV2LET_PREPROC_HPP */
\ No newline at end of file
+#endif /* KWS_ASR_WAV2LETTER_PREPROCESS_HPP */
\ No newline at end of file