diff options
author | George Gekov <george.gekov@arm.com> | 2021-08-16 11:32:10 +0100 |
---|---|---|
committer | Jim Flynn <jim.flynn@arm.com> | 2022-02-05 19:49:06 +0000 |
commit | 23c26277086c78704a17f0dae86da947816320c0 (patch) | |
tree | 88b02fd1fae3130256d059251788a7ef68d2831f /samples/SpeechRecognition | |
parent | 922b912fd2d462bac0809bac5669310ad1506310 (diff) | |
download | armnn-23c26277086c78704a17f0dae86da947816320c0.tar.gz |
MLECO-2079 Adding the C++ KWS example
Signed-off-by: Eanna O Cathain <eanna.ocathain@arm.com>
Change-Id: I81899bbfaada32f478c2e2fc6441eabb94d8d0fc
Diffstat (limited to 'samples/SpeechRecognition')
23 files changed, 648 insertions, 1752 deletions
diff --git a/samples/SpeechRecognition/CMakeLists.txt b/samples/SpeechRecognition/CMakeLists.txt index 6c6b0b6dfc..296a2511dd 100644 --- a/samples/SpeechRecognition/CMakeLists.txt +++ b/samples/SpeechRecognition/CMakeLists.txt @@ -1,4 +1,4 @@ -# Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +# Copyright © 2021 Arm Ltd and Contributors. All rights reserved. # SPDX-License-Identifier: MIT cmake_minimum_required(VERSION 3.0.2) @@ -43,9 +43,11 @@ include(../common/cmake/find_armnn.cmake) include_directories(include) include_directories(../common/include/ArmnnUtils) include_directories(../common/include/Utils) +include_directories(../common/include/Audio) file(GLOB SOURCES "src/*.cpp") file(GLOB COMMON_UTILS_SOURCES "../common/src/Utils/*.cpp") +file(GLOB COMMON_AUDIO_SOURCES "../common/src/Audio/*.cpp") list(REMOVE_ITEM SOURCES ${CMAKE_CURRENT_SOURCE_DIR}/src/Main.cpp) file(GLOB TEST_SOURCES "test/*.cpp") file(GLOB APP_MAIN "src/Main.cpp") @@ -56,7 +58,7 @@ endif() set(APP_TARGET_NAME "${CMAKE_PROJECT_NAME}") -add_executable("${APP_TARGET_NAME}" ${COMMON_UTILS_SOURCES} ${SOURCES} ${APP_MAIN}) +add_executable("${APP_TARGET_NAME}" ${COMMON_UTILS_SOURCES} ${COMMON_AUDIO_SOURCES} ${SOURCES} ${APP_MAIN}) target_link_libraries("${APP_TARGET_NAME}" PUBLIC ${ARMNN_LIBS} -lsndfile -lsamplerate) target_include_directories("${APP_TARGET_NAME}" PUBLIC ${ARMNN_INCLUDE_DIR} ) diff --git a/samples/SpeechRecognition/cmake/unit_tests.cmake b/samples/SpeechRecognition/cmake/unit_tests.cmake index 47c4f4b579..955eed4510 100644 --- a/samples/SpeechRecognition/cmake/unit_tests.cmake +++ b/samples/SpeechRecognition/cmake/unit_tests.cmake @@ -1,4 +1,4 @@ -# Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +# Copyright © 2021 Arm Ltd and Contributors. All rights reserved. # SPDX-License-Identifier: MIT set(TEST_RESOURCES_DIR ${CMAKE_SOURCE_DIR}/test/resources) @@ -7,11 +7,12 @@ add_definitions (-DTEST_RESOURCE_DIR="${TEST_RESOURCES_DIR}") set(TEST_TARGET_NAME "${CMAKE_PROJECT_NAME}-tests") file(GLOB TEST_SOURCES "test/*") +file(GLOB TESTS_AUDIO_COMMON "../common/test/Audio/*") file(MAKE_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/test/resources) include(../common/cmake/find_catch.cmake) -add_executable("${TEST_TARGET_NAME}" ${COMMON_UTILS_SOURCES} ${SOURCES} ${TEST_SOURCES} ) +add_executable("${TEST_TARGET_NAME}" ${COMMON_UTILS_SOURCES} ${COMMON_AUDIO_SOURCES} ${SOURCES} ${TEST_SOURCES} ${TESTS_AUDIO_COMMON}) ExternalProject_Add(passport URL https://raw.githubusercontent.com/Azure-Samples/cognitive-services-speech-sdk/master/sampledata/audiofiles/myVoiceIsMyPassportVerifyMe04.wav diff --git a/samples/SpeechRecognition/include/AudioCapture.hpp b/samples/SpeechRecognition/include/AudioCapture.hpp deleted file mode 100644 index 90c2eccacf..0000000000 --- a/samples/SpeechRecognition/include/AudioCapture.hpp +++ /dev/null @@ -1,62 +0,0 @@ -// -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <string> -#include <iostream> - -#include <math.h> - -#include <vector> - -#include <exception> - -#include "SlidingWindow.hpp" - -namespace asr -{ - -/** -* @brief Class used to capture the audio data loaded from file, and to provide a method of - * extracting correctly positioned and appropriately sized audio windows -* -*/ - class AudioCapture - { - public: - - SlidingWindow<const float> m_window; - int lastReadIdx= 0; - - /** - * @brief Default constructor - */ - AudioCapture() - {}; - - /** - * @brief Function to load the audio data captured from the - * input file to memory. - */ - std::vector<float> LoadAudioFile(std::string filePath); - - /** - * @brief Function to initialize the sliding window. This will set its position in memory, its - * window size and its stride. - */ - void InitSlidingWindow(float* data, size_t dataSize, int minSamples, size_t stride); - - /** - * Checks whether there is another block of audio in memory to read - */ - bool HasNext(); - - /** - * Retrieves the next block of audio if its available - */ - std::vector<float> Next(); - }; -} // namespace asr
\ No newline at end of file diff --git a/samples/SpeechRecognition/include/DataStructures.hpp b/samples/SpeechRecognition/include/DataStructures.hpp deleted file mode 100644 index 9922265299..0000000000 --- a/samples/SpeechRecognition/include/DataStructures.hpp +++ /dev/null @@ -1,102 +0,0 @@ -// -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// -#pragma once - -#include <stdio.h> -#include <iterator> - -/** - * 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 -{ -private: - size_t m_rows; - size_t m_cols; - T* m_data; - -public: - /** - * Creates the array2d with the given sizes. - * - * @param rows number of rows. - * @param cols number of columns. - */ - Array2d(unsigned rows, unsigned cols) - { - if (rows == 0 || cols == 0) { - printf("Array2d constructor has 0 size.\n"); - m_data = nullptr; - return; - } - m_rows = rows; - m_cols = cols; - m_data = new T[rows * cols]; - } - - ~Array2d() - { - delete[] m_data; - } - - T& operator() (unsigned int row, unsigned int col) - { - return m_data[m_cols * row + col]; - } - - T operator() (unsigned int row, unsigned int col) const - { - return m_data[m_cols * row + col]; - } - - /** - * 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; - } - } - - /** - * 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(); }; -}; diff --git a/samples/SpeechRecognition/include/Decoder.hpp b/samples/SpeechRecognition/include/Decoder.hpp index 69d97ccf64..9dd484a5d1 100644 --- a/samples/SpeechRecognition/include/Decoder.hpp +++ b/samples/SpeechRecognition/include/Decoder.hpp @@ -46,8 +46,8 @@ namespace asr rowVector.emplace_back(static_cast<int16_t>(contextToProcess[row * rowLength + j])); } - int max_index = std::distance(rowVector.begin(),std::max_element(rowVector.begin(), rowVector.end())); - unfilteredText.emplace_back(this->m_labels.at(max_index)[0]); + int maxIndex = std::distance(rowVector.begin(), std::max_element(rowVector.begin(), rowVector.end())); + unfilteredText.emplace_back(this->m_labels.at(maxIndex)[0]); } std::string filteredText = FilterCharacters(unfilteredText); diff --git a/samples/SpeechRecognition/include/MFCC.hpp b/samples/SpeechRecognition/include/MFCC.hpp deleted file mode 100644 index 14b6d9fe79..0000000000 --- a/samples/SpeechRecognition/include/MFCC.hpp +++ /dev/null @@ -1,244 +0,0 @@ -// -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <vector> -#include <cstdint> -#include <cmath> -#include <limits> -#include <string> - -/* MFCC's consolidated parameters */ -class MfccParams -{ -public: - float m_samplingFreq; - int m_numFbankBins; - float m_melLoFreq; - float m_melHiFreq; - int m_numMfccFeatures; - int m_frameLen; - int m_frameLenPadded; - bool m_useHtkMethod; - int m_numMfccVectors; - - /** @brief Constructor */ - MfccParams(const float samplingFreq, const int numFbankBins, - const float melLoFreq, const float melHiFreq, - const int numMfccFeats, const int frameLen, - const bool useHtkMethod, const int numMfccVectors); - - /* Delete the default constructor */ - MfccParams() = delete; - - /* Default destructor */ - ~MfccParams() = default; - - /** @brief String representation of parameters */ - std::string Str(); -}; - -/** - * @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 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<float>& audioData); - - MfccParams _m_params; - - /** - * @brief Constructor - * @param[in] params - MFCC parameters - */ - MFCC(const MfccParams& params); - - /* Delete the default constructor */ - MFCC() = delete; - - /** @brief Default destructor */ - ~MFCC() = default; - - /** @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<float>& 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 logStep = 1.8562979903656 / 27.0; - static constexpr float freqStep = 200.0 / 3; - static constexpr float minLogHz = 1000.0; - static constexpr float minLogMel = minLogHz / 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] freq - 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<int32_t>& filterBankFilterFirst, - std::vector<int32_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( - const int32_t inputLength, - const 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 - */ - virtual float GetMelFilterBankNormaliser( - const float& leftMel, - const float& rightMel, - const bool useHTKMethod); - -private: - - 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<int32_t> _m_filterBankFilterFirst; - std::vector<int32_t> _m_filterBankFilterLast; - bool _m_filterBankInitialised; - - /** - * @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(); - - /** - * @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<float>& audioData); - - /** @brief Computes the magnitude from an interleaved complex array */ - void _ConvertToPowerSpectrum(); - -}; - diff --git a/samples/SpeechRecognition/include/MathUtils.hpp b/samples/SpeechRecognition/include/MathUtils.hpp deleted file mode 100644 index 5f81fb6507..0000000000 --- a/samples/SpeechRecognition/include/MathUtils.hpp +++ /dev/null @@ -1,85 +0,0 @@ -// -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include <vector> -#include <cmath> -#include <cstdint> -#include <numeric> - -class MathUtils -{ - -public: - - /** - * @brief Computes the FFT for the input vector - * @param[in] input Floating point vector of input elements - * @param[out] fftOutput Output buffer to be populated by computed - * FFTs - * @return none - */ - static void FftF32(std::vector<float>& input, - std::vector<float>& fftOutput); - - - /** - * @brief Computes the dot product of two 1D floating point - * vectors. - * result = sum(srcA[0]*srcB[0] + srcA[1]*srcB[1] + ..) - * @param[in] srcPtrA pointer to the first element of first - * array - * @param[in] srcPtrB pointer to the first element of second - * array - * @param[in] srcLen Number of elements in the array/vector - * @return dot product - */ - static float DotProductF32(float* srcPtrA, float* srcPtrB, - const int srcLen); - - /** - * @brief Computes the squared magnitude of floating point - * complex number array. - * @param[in] ptrSrc pointer to the first element of input - * array - * @param[in] srcLen Number of elements in the array/vector - * @param[out] ptrDst Output buffer to be populated - * @param[in] dstLen output buffer len (for sanity check only) - * @return true if successful, false otherwise - */ - static bool ComplexMagnitudeSquaredF32(float* ptrSrc, - const int srcLen, - float* ptrDst, - const int dstLen); - - /** - * @brief Computes the natural logarithms of input floating point - * vector - * @param[in] input Floating point input vector - * @param[out] output Pre-allocated buffer to be populated with - * natural log values of each input element - * @return none - */ - static void VecLogarithmF32(std::vector <float>& input, - std::vector <float>& output); - - /** - * @brief Gets the mean of a floating point array of elements - * @param[in] ptrSrc pointer to the first element - * @param[in] srcLen Number of elements in the array/vector - * @return average value - */ - static float MeanF32(float* ptrSrc, const uint32_t srcLen); - - /** - * @brief Gets the standard deviation of a floating point array - * of elements - * @param[in] ptrSrc pointer to the first element - * @param[in] srcLen Number of elements in the array/vector - * @param[in] mean pre-computed mean value - * @return standard deviation value - */ - static float StdDevF32(float* ptrSrc, const uint32_t srcLen, - const float mean); -}; diff --git a/samples/SpeechRecognition/include/SlidingWindow.hpp b/samples/SpeechRecognition/include/SlidingWindow.hpp deleted file mode 100644 index 791a0b7fc0..0000000000 --- a/samples/SpeechRecognition/include/SlidingWindow.hpp +++ /dev/null @@ -1,161 +0,0 @@ -// -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -template<class T> -class SlidingWindow -{ -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; -public: - - /** - * Creates the window slider through the given data. - * - * @param data pointer to the data to slide through. - * @param dataSize size in T type elements wise. - * @param windowSize sliding window size in T type wise elements. - * @param 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; - - /** - * 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; - } - } - - /** - * 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); - } - - /** - * Resest the slider to the initial position. - */ - virtual void Reset() - { - m_count = 0; - } - - /** - * Resest the slider to the initial position. - */ - virtual size_t GetWindowSize() - { - return m_size; - } - - /** - * Resets the slider to the start of the new data. - * New data size MUST be the same as the old one. - * @param newStart pointer to the new data to slide through. - */ - virtual void Reset(T* newStart) - { - m_start = newStart; - Reset(); - } - - /** - * 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; - } - - /** - * 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 size_t NextWindowStartIndex() - { - return m_count == 0? 0: ((m_count) * m_stride); - } - - /** - * Go to given sliding window index. - * @param 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; - } - - /** - * Calculates whole number of times the window can stride through the given data. - * @return maximum number of strides. - */ - size_t TotalStrides() - { - if (m_size > m_dataSize) - { - return 0; - } - return ((m_dataSize - m_size)/m_stride); - } - - /** - * 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_size > this->m_dataSize) - { - return this->m_dataSize / this->m_size; - } - else - { - return ((this->m_dataSize - this->m_size)/ static_cast<float>(this->m_stride)); - } - - } - - /** - * Calculates the remaining data left to be processed - * @return The remaining unprocessed data - */ - int RemainingData() - { - return this->m_dataSize - this->NextWindowStartIndex(); - } -};
\ No newline at end of file diff --git a/samples/SpeechRecognition/include/SpeechRecognitionPipeline.hpp b/samples/SpeechRecognition/include/SpeechRecognitionPipeline.hpp index 47ce30416f..bc3fbfe151 100644 --- a/samples/SpeechRecognition/include/SpeechRecognitionPipeline.hpp +++ b/samples/SpeechRecognition/include/SpeechRecognitionPipeline.hpp @@ -8,16 +8,16 @@ #include "ArmnnNetworkExecutor.hpp" #include "Decoder.hpp" #include "MFCC.hpp" -#include "Preprocess.hpp" +#include "Wav2LetterPreprocessor.hpp" -namespace asr +namespace asr { /** * Generic Speech Recognition pipeline with 3 steps: data pre-processing, inference execution and inference * result post-processing. * */ -class ASRPipeline +class ASRPipeline { public: @@ -27,7 +27,7 @@ public: * @param decoder - unique pointer to inference results decoder */ ASRPipeline(std::unique_ptr<common::ArmnnNetworkExecutor<int8_t>> executor, - std::unique_ptr<Decoder> decoder); + std::unique_ptr<Decoder> decoder, std::unique_ptr<Wav2LetterPreprocessor> preprocessor); /** * @brief Standard audio pre-processing implementation. @@ -36,20 +36,16 @@ public: * extracting the MFCC features. * @param[in] audio - the raw audio data - * @param[out] preprocessor - the preprocessor object, which handles the data prepreration + * @param[out] preprocessor - the preprocessor object, which handles the data preparation */ - template<typename Tin,typename Tout> - std::vector<Tout> PreProcessing(std::vector<Tin>& audio, Preprocess& preprocessor) - { - int audioDataToPreProcess = preprocessor._m_windowLen + - ((preprocessor._m_mfcc._m_params.m_numMfccVectors -1) *preprocessor._m_windowStride); - int outputBufferSize = preprocessor._m_mfcc._m_params.m_numMfccVectors - * preprocessor._m_mfcc._m_params.m_numMfccFeatures * 3; - std::vector<Tout> outputBuffer(outputBufferSize); - preprocessor.Invoke(audio.data(), audioDataToPreProcess, outputBuffer, m_executor->GetQuantizationOffset(), - m_executor->GetQuantizationScale()); - return outputBuffer; - } + std::vector<int8_t> PreProcessing(std::vector<float>& audio); + + int getInputSamplesSize(); + int getSlidingWindowOffset(); + + // Exposing hardcoded constant as it can only be derived from model knowledge and not from model itself + // Will need to be refactored so that hard coded values are not defined outside of model settings + int SLIDING_WINDOW_OFFSET; /** * @brief Executes inference @@ -60,9 +56,9 @@ public: * @param[out] result - raw inference results. */ template<typename T> - void Inference(const std::vector<T>& preprocessedData, common::InferenceResults<int8_t>& result) + void Inference(const std::vector<T>& preprocessedData, common::InferenceResults<int8_t>& result) { - size_t data_bytes = sizeof(std::vector<T>) + (sizeof(T) * preprocessedData.size()); + size_t data_bytes = sizeof(T) * preprocessedData.size(); m_executor->Run(preprocessedData.data(), data_bytes, result); } @@ -78,9 +74,9 @@ public: */ template<typename T> void PostProcessing(common::InferenceResults<int8_t>& inferenceResult, - bool& isFirstWindow, - bool isLastWindow, - std::string currentRContext) + bool& isFirstWindow, + bool isLastWindow, + std::string currentRContext) { int rowLength = 29; int middleContextStart = 49; @@ -92,17 +88,17 @@ public: std::vector<T> contextToProcess; // If isFirstWindow we keep the left context of the output - if(isFirstWindow) + if (isFirstWindow) { std::vector<T> chunk(&inferenceResult[0][leftContextStart], - &inferenceResult[0][middleContextEnd * rowLength]); + &inferenceResult[0][middleContextEnd * rowLength]); contextToProcess = chunk; } - // Else we only keep the middle context of the output - else + else { + // Else we only keep the middle context of the output std::vector<T> chunk(&inferenceResult[0][middleContextStart * rowLength], - &inferenceResult[0][middleContextEnd * rowLength]); + &inferenceResult[0][middleContextEnd * rowLength]); contextToProcess = chunk; } std::string output = this->m_decoder->DecodeOutput<T>(contextToProcess); @@ -110,10 +106,10 @@ public: std::cout << output << std::flush; // If this is the last window, we print the right context of the output - if(isLastWindow) + if (isLastWindow) { - std::vector<T> rContext(&inferenceResult[0][rightContextStart*rowLength], - &inferenceResult[0][rightContextEnd * rowLength]); + std::vector<T> rContext(&inferenceResult[0][rightContextStart * rowLength], + &inferenceResult[0][rightContextEnd * rowLength]); currentRContext = this->m_decoder->DecodeOutput(rContext); std::cout << currentRContext << std::endl; } @@ -122,6 +118,7 @@ public: protected: std::unique_ptr<common::ArmnnNetworkExecutor<int8_t>> m_executor; std::unique_ptr<Decoder> m_decoder; + std::unique_ptr<Wav2LetterPreprocessor> m_preProcessor; }; using IPipelinePtr = std::unique_ptr<asr::ASRPipeline>; @@ -136,4 +133,4 @@ using IPipelinePtr = std::unique_ptr<asr::ASRPipeline>; */ IPipelinePtr CreatePipeline(common::PipelineOptions& config, std::map<int, std::string>& labels); -}// namespace asr
\ No newline at end of file +} // namespace asr
\ No newline at end of file diff --git a/samples/SpeechRecognition/include/Wav2LetterMFCC.hpp b/samples/SpeechRecognition/include/Wav2LetterMFCC.hpp new file mode 100644 index 0000000000..aa88aafb3b --- /dev/null +++ b/samples/SpeechRecognition/include/Wav2LetterMFCC.hpp @@ -0,0 +1,78 @@ +// +// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// +#pragma once + +#include "MFCC.hpp" + +/* Class to provide Wav2Letter specific MFCC calculation requirements. */ +class Wav2LetterMFCC : public MFCC +{ + +public: + explicit Wav2LetterMFCC(const MfccParams& params) + : MFCC(params) + {} + + 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; +};
\ No newline at end of file diff --git a/samples/SpeechRecognition/include/Preprocess.hpp b/samples/SpeechRecognition/include/Wav2LetterPreprocessor.hpp index 80c568439b..ebc9e864e3 100644 --- a/samples/SpeechRecognition/include/Preprocess.hpp +++ b/samples/SpeechRecognition/include/Wav2LetterPreprocessor.hpp @@ -1,48 +1,23 @@ // -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // +#ifndef SPEECH_RECOGNITION_EXAMPLE_WAV2LETTERPREPROCESSOR_HPP +#define SPEECH_RECOGNITION_EXAMPLE_WAV2LETTERPREPROCESSOR_HPP -#pragma once - +#include <numeric> #include "DataStructures.hpp" #include "SlidingWindow.hpp" -#include <numeric> #include "MFCC.hpp" +#include "Wav2LetterMFCC.hpp" +// Class to facilitate pre-processing calculation for Wav2Letter model for ASR +using AudioWindow = SlidingWindow<const float>; -/* Class to facilitate pre-processing calculation for Wav2Letter model - * for ASR */ -using AudioWindow = SlidingWindow <const float>; - -class Preprocess +class Wav2LetterPreprocessor { public: - - MFCC _m_mfcc; /* MFCC instance */ - - /* 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_windowLen; /* Window length for MFCC */ - uint32_t _m_windowStride; /* Window stride len for MFCC */ - AudioWindow _m_window; /* Sliding window */ - - /** - * @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( - const uint32_t windowLen, - const uint32_t windowStride, - const MFCC mfccInst); - Preprocess() = delete; - ~Preprocess(); + Wav2LetterPreprocessor(uint32_t windowLen, uint32_t windowStride, + std::unique_ptr<Wav2LetterMFCC> mfccInst); /** * @brief Calculates the features required from audio data. This @@ -55,12 +30,19 @@ public: * @param[in] tensor tensor to be populated * @return true if successful, false in case of error. */ - bool Invoke(const float* audioData, - const uint32_t audioDataLen, - std::vector<int8_t>& output, - int quantOffset, + bool Invoke(const float* audioData, uint32_t audioDataLen, std::vector<int8_t>& output, int quantOffset, float quantScale); + std::unique_ptr<MFCC> m_mfcc; + + // 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_windowLen; // Window length for MFCC + uint32_t m_windowStride; // Window stride len for MFCC + AudioWindow m_window; // Sliding window protected: /** @@ -73,16 +55,18 @@ protected: * * @return true if successful, false otherwise */ - static bool _ComputeDeltas(Array2d<float>& mfcc, - Array2d<float>& delta1, - Array2d<float>& delta2); + static bool ComputeDeltas(Array2d<float>& mfcc, + Array2d<float>& delta1, + Array2d<float>& delta2); + +protected: /** * @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); + static float GetMean(Array2d<float>& vec); /** * @brief Given a 2D vector of floats, computes the stddev @@ -90,8 +74,7 @@ protected: * @param[in] mean mean value of the vector passed in * @return stddev value */ - static float _GetStdDev(Array2d<float>& vec, - const float mean); + static float GetStdDev(Array2d<float>& vec, float mean); /** * @brief Given a 2D vector of floats, normalises it using @@ -99,13 +82,13 @@ protected: * @param[in/out] vec vector of vector of floats * @return */ - static void _NormaliseVec(Array2d<float>& vec); + static void NormaliseVec(Array2d<float>& vec); /** * @brief Normalises the MFCC and delta buffers * @return */ - void _Normalise(); + void Normalise(); /** * @brief Given the quantisation and data type limits, computes @@ -117,12 +100,12 @@ protected: * @param[in] maxVal Numerical limit - maximum * @return floating point quantised value */ - static float _GetQuantElem( - const float elem, - const float quantScale, - const int quantOffset, - const float minVal, - const float maxVal); + static float GetQuantElem( + float elem, + float quantScale, + int quantOffset, + float minVal, + float maxVal); /** * @brief Quantises the MFCC and delta buffers, and places them @@ -137,39 +120,39 @@ protected: * @param[in] quantScale quantisation scale * @param[in] quantOffset quantisation offset */ - template <typename T> - bool _Quantise(T* outputBuf, int quantOffset, float quantScale) + template<typename T> + bool Quantise(T*outputBuf, int quantOffset, float quantScale) { - /* Populate */ + // Populate T* outputBufMfcc = outputBuf; - T* outputBufD1 = outputBuf + this->_m_mfcc._m_params.m_numMfccFeatures; - T* outputBufD2 = outputBufD1 + this->_m_mfcc._m_params.m_numMfccFeatures; - const uint32_t ptrIncr = this->_m_mfcc._m_params.m_numMfccFeatures * 2; /* (3 vectors - 1 vector) */ + T* outputBufD1 = outputBuf + this->m_mfcc->m_params.m_numMfccFeatures; + T* outputBufD2 = outputBufD1 + this->m_mfcc->m_params.m_numMfccFeatures; + const uint32_t ptrIncr = this->m_mfcc->m_params.m_numMfccFeatures * 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_mfcc._m_params.m_numMfccVectors; ++j) { - for (uint32_t i = 0; i < this->_m_mfcc._m_params.m_numMfccFeatures; ++i) + // We need to do a transpose while copying and concatenating the tensor + for (uint32_t j = 0; j < this->m_mfcc->m_params.m_numMfccVectors; ++j) + { + for (uint32_t i = 0; i < this->m_mfcc->m_params.m_numMfccFeatures; ++i) { - *outputBufMfcc++ = static_cast<T>(this->_GetQuantElem( - this->_m_mfccBuf(i, j), quantScale, + *outputBufMfcc++ = static_cast<T>(Wav2LetterPreprocessor::GetQuantElem( + this->m_mfccBuf(i, j), quantScale, quantOffset, minVal, maxVal)); - *outputBufD1++ = static_cast<T>(this->_GetQuantElem( - this->_m_delta1Buf(i, j), quantScale, + *outputBufD1++ = static_cast<T>(Wav2LetterPreprocessor::GetQuantElem( + this->m_delta1Buf(i, j), quantScale, quantOffset, minVal, maxVal)); - *outputBufD2++ = static_cast<T>(this->_GetQuantElem( - this->_m_delta2Buf(i, j), quantScale, + *outputBufD2++ = static_cast<T>(Wav2LetterPreprocessor::GetQuantElem( + this->m_delta2Buf(i, j), quantScale, quantOffset, minVal, maxVal)); } outputBufMfcc += ptrIncr; outputBufD1 += ptrIncr; outputBufD2 += ptrIncr; } - return true; } }; +#endif //SPEECH_RECOGNITION_EXAMPLE_WAV2LETTERPREPROCESSOR_HPP diff --git a/samples/SpeechRecognition/src/AudioCapture.cpp b/samples/SpeechRecognition/src/AudioCapture.cpp deleted file mode 100644 index f3b9092218..0000000000 --- a/samples/SpeechRecognition/src/AudioCapture.cpp +++ /dev/null @@ -1,104 +0,0 @@ -// -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "AudioCapture.hpp" -#include <alsa/asoundlib.h> -#include <sndfile.h> -#include <samplerate.h> - -namespace asr -{ - std::vector<float> AudioCapture::LoadAudioFile(std::string filePath) - { - SF_INFO inputSoundFileInfo; - SNDFILE* infile = NULL; - infile = sf_open(filePath.c_str(), SFM_READ, &inputSoundFileInfo); - - float audioIn[inputSoundFileInfo.channels * inputSoundFileInfo.frames]; - sf_read_float(infile, audioIn, inputSoundFileInfo.channels * inputSoundFileInfo.frames); - - float sampleRate = 16000.0f; - float srcRatio = sampleRate / (float)inputSoundFileInfo.samplerate; - int outputFrames = ceil(inputSoundFileInfo.frames * srcRatio); - float dataOut[outputFrames]; - - // Convert to mono - float monoData[inputSoundFileInfo.frames]; - for(int i = 0; i < inputSoundFileInfo.frames; i++) - { - float val = 0.0f; - for(int j = 0; j < inputSoundFileInfo.channels; j++) - monoData[i] += audioIn[i * inputSoundFileInfo.channels + j]; - monoData[i] /= inputSoundFileInfo.channels; - } - - // Resample - SRC_DATA srcData; - srcData.data_in = monoData; - srcData.input_frames = inputSoundFileInfo.frames; - srcData.data_out = dataOut; - srcData.output_frames = outputFrames; - srcData.src_ratio = srcRatio; - - src_simple(&srcData, SRC_SINC_BEST_QUALITY, 1); - - // Convert to Vector - std::vector<float> processedInput; - - for(int i = 0; i < srcData.output_frames_gen; ++i) - { - processedInput.push_back(srcData.data_out[i]); - } - - sf_close(infile); - - return processedInput; - } - - void AudioCapture::InitSlidingWindow(float* data, size_t dataSize, int minSamples, size_t stride) - { - this->m_window = SlidingWindow<const float>(data, dataSize, minSamples, stride); - } - - bool AudioCapture::HasNext() - { - return m_window.HasNext(); - } - - std::vector<float> AudioCapture::Next() - { - if (this->m_window.HasNext()) - { - int remainingData = this->m_window.RemainingData(); - const float* windowData = this->m_window.Next(); - - size_t windowSize = this->m_window.GetWindowSize(); - - if(remainingData < windowSize) - { - std::vector<float> mfccAudioData(windowSize, 0.0f); - for(int i = 0; i < remainingData; ++i) - { - mfccAudioData[i] = *windowData; - if(i < remainingData - 1) - { - ++windowData; - } - } - return mfccAudioData; - } - else - { - std::vector<float> mfccAudioData(windowData, windowData + windowSize); - return mfccAudioData; - } - } - else - { - throw std::out_of_range("Error, end of audio data reached."); - } - } -} //namespace asr - diff --git a/samples/SpeechRecognition/src/Decoder.cpp b/samples/SpeechRecognition/src/Decoder.cpp index 663d4db5b5..b95288e95c 100644 --- a/samples/SpeechRecognition/src/Decoder.cpp +++ b/samples/SpeechRecognition/src/Decoder.cpp @@ -5,33 +5,32 @@ #include "Decoder.hpp" -namespace asr { +namespace asr +{ - Decoder::Decoder(std::map<int, std::string>& labels): - m_labels(labels) - {} +Decoder::Decoder(std::map<int, std::string>& labels) : + m_labels(labels) {} - std::string Decoder::FilterCharacters(std::vector<char>& unfiltered) - { - std::string filtered = ""; +std::string Decoder::FilterCharacters(std::vector<char>& unfiltered) +{ + std::string filtered; - for(int i = 0; i < unfiltered.size(); ++i) + for (int i = 0; i < unfiltered.size(); ++i) + { + if (unfiltered.at(i) == '$') { - if (unfiltered.at(i) == '$') - { - continue; - } - - else if (i + 1 < unfiltered.size() && unfiltered.at(i) == unfiltered.at(i + 1)) - { - continue; - } - else - { - filtered += unfiltered.at(i); - } + continue; + } + else if (i + 1 < unfiltered.size() && unfiltered.at(i) == unfiltered.at(i + 1)) + { + continue; + } + else + { + filtered += unfiltered.at(i); } - return filtered; } -}// namespace + return filtered; +} +} // namespace asr diff --git a/samples/SpeechRecognition/src/MFCC.cpp b/samples/SpeechRecognition/src/MFCC.cpp deleted file mode 100644 index 234b14d3be..0000000000 --- a/samples/SpeechRecognition/src/MFCC.cpp +++ /dev/null @@ -1,397 +0,0 @@ -// -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include <cstdio> -#include <float.h> - -#include "MFCC.hpp" -#include "MathUtils.hpp" - - -MfccParams::MfccParams( - const float samplingFreq, - const int numFbankBins, - const float melLoFreq, - const float melHiFreq, - const int numMfccFeats, - const int frameLen, - const bool useHtkMethod, - const int numMfccVectors): - m_samplingFreq(samplingFreq), - m_numFbankBins(numFbankBins), - m_melLoFreq(melLoFreq), - m_melHiFreq(melHiFreq), - m_numMfccFeatures(numMfccFeats), - m_frameLen(frameLen), - m_numMfccVectors(numMfccVectors), - - /* Smallest power of 2 >= frame length. */ - m_frameLenPadded(pow(2, ceil((log(frameLen)/log(2))))), - m_useHtkMethod(useHtkMethod) -{} - -std::string MfccParams::Str() -{ - char strC[1024]; - snprintf(strC, sizeof(strC) - 1, "\n \ - \n\t Sampling frequency: %f\ - \n\t Number of filter banks: %u\ - \n\t Mel frequency limit (low): %f\ - \n\t Mel frequency limit (high): %f\ - \n\t Number of MFCC features: %u\ - \n\t Frame length: %u\ - \n\t Padded frame length: %u\ - \n\t Using HTK for Mel scale: %s\n", - this->m_samplingFreq, this->m_numFbankBins, this->m_melLoFreq, - this->m_melHiFreq, this->m_numMfccFeatures, this->m_frameLen, - this->m_frameLenPadded, this->m_useHtkMethod ? "yes" : "no"); - return std::string{strC}; -} - -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 float multiplier = 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 * cos(static_cast<float>(i) * multiplier))); - } -} - -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 / freqStep; - - if (freq >= minLogHz) - { - mel = minLogMel + logf(freq / minLogHz) / 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 = freqStep * melFreq; - - if (melFreq >= minLogMel) - { - freq = minLogHz * expf(logStep * (melFreq - minLogMel)); - } - return freq; - } -} - - -bool MFCC::ApplyMelFilterBank( - std::vector<float>& fftVec, - std::vector<std::vector<float>>& melFilterBank, - std::vector<int32_t>& filterBankFilterFirst, - std::vector<int32_t>& filterBankFilterLast, - std::vector<float>& melEnergies) -{ - const size_t numBanks = melEnergies.size(); - - if (numBanks != filterBankFilterFirst.size() || - numBanks != filterBankFilterLast.size()) - { - printf("unexpected filter bank lengths\n"); - return false; - } - - for (size_t bin = 0; bin < numBanks; ++bin) - { - auto filterBankIter = melFilterBank[bin].begin(); - float melEnergy = 1e-10; /* Avoid log of zero at later stages */ - const int32_t firstIndex = filterBankFilterFirst[bin]; - const int32_t lastIndex = filterBankFilterLast[bin]; - - for (int32_t i = firstIndex; i <= lastIndex; ++i) - { - melEnergy += (*filterBankIter++ * fftVec[i]); - } - - melEnergies[bin] = melEnergy; - } - - return true; -} - -void MFCC::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 */ - MathUtils::VecLogarithmF32(melEnergies, vecLogEnergies); - - /* Scale the log values and get the max */ - for (auto iterM = melEnergies.begin(), iterL = vecLogEnergies.begin(); - iterM != melEnergies.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 (auto iter = melEnergies.begin(); iter != melEnergies.end(); ++iter) - { - *iter = std::max(*iter, clampLevelLowdB); - } -} - -void MFCC::_ConvertToPowerSpectrum() -{ - const uint32_t halfDim = this->_m_params.m_frameLenPadded / 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]; - - 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); - - /* Orthonormal normalization. */ - const float normalizerK0 = 2 * sqrt(1.0 / static_cast<float>(4*inputLength)); - const float normalizer = 2 * sqrt(1.0 / 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; - } - - /* 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 * - cos((n + 0.5) * angle); - } - angle += angleIncr; - } - return dctMatix; -} - -float MFCC::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))); -} - -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() -{ - return this->_m_filterBankInitialised; -} - -void MFCC::_MfccComputePreFeature(const std::vector<float>& audioData) -{ - this->_InitMelFilterBank(); - - /* TensorFlow way of normalizing .wav data to (-1, 1). */ - constexpr float normaliser = 1.0; - 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. */ - MathUtils::FftF32(this->_m_frame, this->_m_buffer); - - /* 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("Failed to apply MEL filter banks\n"); - } - - /* Convert to logarithmic scale */ - this->ConvertToLogarithmicScale(this->_m_melEnergies); -} - -std::vector<float> MFCC::MfccCompute(const std::vector<float>& 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++ = 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<int32_t>(this->_m_params.m_numFbankBins); - this->_m_filterBankFilterLast = - std::vector<int32_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; - - int32_t firstIndex = -1; - int32_t lastIndex = -1; - 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 (firstIndex == -1) - { - firstIndex = i; - } - lastIndex = i; - } - } - - this->_m_filterBankFilterFirst[bin] = firstIndex; - this->_m_filterBankFilterLast[bin] = lastIndex; - - /* Copy the part we care about. */ - for (int32_t i = firstIndex; i <= lastIndex; i++) - { - melFilterBank[bin].push_back(thisBin[i]); - } - } - - return melFilterBank; -} - diff --git a/samples/SpeechRecognition/src/Main.cpp b/samples/SpeechRecognition/src/Main.cpp index de37e23b40..e2d293001f 100644 --- a/samples/SpeechRecognition/src/Main.cpp +++ b/samples/SpeechRecognition/src/Main.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2021 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #include <iostream> @@ -11,10 +11,8 @@ #include "CmdArgsParser.hpp" #include "ArmnnNetworkExecutor.hpp" #include "AudioCapture.hpp" -#include "Preprocess.hpp" -#include "Decoder.hpp" #include "SpeechRecognitionPipeline.hpp" - +#include "Wav2LetterMFCC.hpp" using InferenceResult = std::vector<int8_t>; using InferenceResults = std::vector<InferenceResult>; @@ -25,101 +23,77 @@ const std::string LABEL_PATH = "--label-path"; const std::string PREFERRED_BACKENDS = "--preferred-backends"; const std::string HELP = "--help"; -std::map<int, std::string> labels = { - {0, "a" }, - {1, "b" }, - {2, "c" }, - {3, "d" }, - {4, "e" }, - {5, "f" }, - {6, "g" }, - {7, "h" }, - {8, "i" }, - {9, "j" }, - {10,"k" }, - {11,"l" }, - {12,"m" }, - {13,"n" }, - {14,"o" }, - {15,"p" }, - {16,"q" }, - {17,"r" }, - {18,"s" }, - {19,"t" }, - {20,"u" }, - {21,"v" }, - {22,"w" }, - {23,"x" }, - {24,"y" }, - {25,"z" }, - {26, "\'" }, +std::map<int, std::string> labels = +{ + {0, "a"}, + {1, "b"}, + {2, "c"}, + {3, "d"}, + {4, "e"}, + {5, "f"}, + {6, "g"}, + {7, "h"}, + {8, "i"}, + {9, "j"}, + {10, "k"}, + {11, "l"}, + {12, "m"}, + {13, "n"}, + {14, "o"}, + {15, "p"}, + {16, "q"}, + {17, "r"}, + {18, "s"}, + {19, "t"}, + {20, "u"}, + {21, "v"}, + {22, "w"}, + {23, "x"}, + {24, "y"}, + {25, "z"}, + {26, "\'"}, {27, " "}, - {28,"$" } + {28, "$"} }; /* * The accepted options for this Speech Recognition executable */ -static std::map<std::string, std::string> CMD_OPTIONS = { - {AUDIO_FILE_PATH, "[REQUIRED] Path to the Audio file to run speech recognition on"}, - {MODEL_FILE_PATH, "[REQUIRED] Path to the Speech Recognition model to use"}, - {PREFERRED_BACKENDS, "[OPTIONAL] Takes the preferred backends in preference order, separated by comma." - " For example: CpuAcc,GpuAcc,CpuRef. Accepted options: [CpuAcc, CpuRef, GpuAcc]." - " Defaults to CpuAcc,CpuRef"} +static std::map<std::string, std::string> CMD_OPTIONS = +{ + {AUDIO_FILE_PATH, "[REQUIRED] Path to the Audio file to run speech recognition on"}, + {MODEL_FILE_PATH, "[REQUIRED] Path to the Speech Recognition model to use"}, + {PREFERRED_BACKENDS, "[OPTIONAL] Takes the preferred backends in preference order, separated by comma." + " For example: CpuAcc,GpuAcc,CpuRef. Accepted options: [CpuAcc, CpuRef, GpuAcc]." + " Defaults to CpuAcc,CpuRef"} }; /* * Reads the user supplied backend preference, splits it by comma, and returns an ordered vector */ -std::vector<armnn::BackendId> GetPreferredBackendList(const std::string& preferredBackends) +std::vector<armnn::BackendId> GetPreferredBackendList(const std::string& preferredBackends) { std::vector<armnn::BackendId> backends; std::stringstream ss(preferredBackends); - while(ss.good()) + while (ss.good()) { std::string backend; - std::getline( ss, backend, ',' ); + std::getline(ss, backend, ','); backends.emplace_back(backend); } return backends; } -int main(int argc, char *argv[]) +int main(int argc, char* argv[]) { - // Wav2Letter ASR SETTINGS - int SAMP_FREQ = 16000; - int FRAME_LEN_MS = 32; - int FRAME_LEN_SAMPLES = SAMP_FREQ * FRAME_LEN_MS * 0.001; - int NUM_MFCC_FEATS = 13; - int MFCC_WINDOW_LEN = 512; - int MFCC_WINDOW_STRIDE = 160; - const int NUM_MFCC_VECTORS = 296; - int SAMPLES_PER_INFERENCE = MFCC_WINDOW_LEN + ((NUM_MFCC_VECTORS -1) * MFCC_WINDOW_STRIDE); - int MEL_LO_FREQ = 0; - int MEL_HI_FREQ = 8000; - int NUM_FBANK_BIN = 128; - int INPUT_WINDOW_LEFT_CONTEXT = 98; - int INPUT_WINDOW_RIGHT_CONTEXT = 98; - int INPUT_WINDOW_INNER_CONTEXT = NUM_MFCC_VECTORS - - (INPUT_WINDOW_LEFT_CONTEXT + INPUT_WINDOW_RIGHT_CONTEXT); - int SLIDING_WINDOW_OFFSET = INPUT_WINDOW_INNER_CONTEXT * MFCC_WINDOW_STRIDE; - - - MfccParams mfccParams(SAMP_FREQ, NUM_FBANK_BIN, - MEL_LO_FREQ, MEL_HI_FREQ, NUM_MFCC_FEATS, FRAME_LEN_SAMPLES, false, NUM_MFCC_VECTORS); - - MFCC mfccInst = MFCC(mfccParams); - - Preprocess preprocessor(MFCC_WINDOW_LEN, MFCC_WINDOW_STRIDE, mfccInst); - bool isFirstWindow = true; - std::string currentRContext = ""; + std::string currentRContext = ""; - std::map <std::string, std::string> options; + std::map<std::string, std::string> options; int result = ParseOptions(options, CMD_OPTIONS, argv, argc); - if (result != 0) + if (result != 0) { return result; } @@ -127,28 +101,29 @@ int main(int argc, char *argv[]) // Create the network options common::PipelineOptions pipelineOptions; pipelineOptions.m_ModelFilePath = GetSpecifiedOption(options, MODEL_FILE_PATH); - - if (CheckOptionSpecified(options, PREFERRED_BACKENDS)) + pipelineOptions.m_ModelName = "Wav2Letter"; + if (CheckOptionSpecified(options, PREFERRED_BACKENDS)) { pipelineOptions.m_backends = GetPreferredBackendList((GetSpecifiedOption(options, PREFERRED_BACKENDS))); - } - else + } + else { pipelineOptions.m_backends = {"CpuAcc", "CpuRef"}; } asr::IPipelinePtr asrPipeline = asr::CreatePipeline(pipelineOptions, labels); - asr::AudioCapture capture; - std::vector<float> audioData = capture.LoadAudioFile(GetSpecifiedOption(options, AUDIO_FILE_PATH)); - capture.InitSlidingWindow(audioData.data(), audioData.size(), SAMPLES_PER_INFERENCE, SLIDING_WINDOW_OFFSET); + audio::AudioCapture capture; + std::vector<float> audioData = audio::AudioCapture::LoadAudioFile(GetSpecifiedOption(options, AUDIO_FILE_PATH)); + capture.InitSlidingWindow(audioData.data(), audioData.size(), asrPipeline->getInputSamplesSize(), + asrPipeline->getSlidingWindowOffset()); - while (capture.HasNext()) + while (capture.HasNext()) { std::vector<float> audioBlock = capture.Next(); InferenceResults results; - std::vector<int8_t> preprocessedData = asrPipeline->PreProcessing<float, int8_t>(audioBlock, preprocessor); + std::vector<int8_t> preprocessedData = asrPipeline->PreProcessing(audioBlock); asrPipeline->Inference<int8_t>(preprocessedData, results); asrPipeline->PostProcessing<int8_t>(results, isFirstWindow, !capture.HasNext(), currentRContext); } diff --git a/samples/SpeechRecognition/src/MathUtils.cpp b/samples/SpeechRecognition/src/MathUtils.cpp deleted file mode 100644 index bf9908343a..0000000000 --- a/samples/SpeechRecognition/src/MathUtils.cpp +++ /dev/null @@ -1,112 +0,0 @@ -// -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "MathUtils.hpp" -#include <vector> -#include <cmath> -#include <cstdio> - -void MathUtils::FftF32(std::vector<float>& input, - std::vector<float>& fftOutput) -{ - const int inputLength = input.size(); - - for (int k = 0; k <= inputLength / 2; k++) - { - float sumReal = 0, sumImag = 0; - - for (int t = 0; t < inputLength; t++) - { - float angle = 2 * M_PI * t * k / inputLength; - sumReal += input[t] * cosf(angle); - sumImag += -input[t] * sinf(angle); - } - - /* Arrange output to [real0, realN/2, real1, im1, real2, im2, ...] */ - if (k == 0) - { - fftOutput[0] = sumReal; - } - else if (k == inputLength / 2) - { - fftOutput[1] = sumReal; - } - else - { - fftOutput[k*2] = sumReal; - fftOutput[k*2 + 1] = sumImag; - }; - } -} - -float MathUtils::DotProductF32(float* srcPtrA, float* srcPtrB, - const int srcLen) -{ - float output = 0.f; - - for (int i = 0; i < srcLen; ++i) - { - output += *srcPtrA++ * *srcPtrB++; - } - return output; -} - -bool MathUtils::ComplexMagnitudeSquaredF32(float* ptrSrc, - const int srcLen, - float* ptrDst, - const int dstLen) -{ - if (dstLen < srcLen/2) - { - printf("dstLen must be greater than srcLen/2"); - return false; - } - - for (int j = 0; j < srcLen; ++j) - { - const float real = *ptrSrc++; - const float im = *ptrSrc++; - *ptrDst++ = real*real + im*im; - } - return true; -} - -void MathUtils::VecLogarithmF32(std::vector <float>& input, - std::vector <float>& output) -{ - for (auto in = input.begin(), out = output.begin(); - in != input.end(); ++in, ++out) - { - *out = logf(*in); - } -} - -float MathUtils::MeanF32(float* ptrSrc, const uint32_t srcLen) -{ - if (!srcLen) - { - return 0.f; - } - - float acc = std::accumulate(ptrSrc, ptrSrc + srcLen, 0.0); - return acc/srcLen; -} - -float MathUtils::StdDevF32(float* ptrSrc, const uint32_t srcLen, - const float mean) -{ - if (!srcLen) - { - return 0.f; - } - auto VarianceFunction = [=](float acc, const float value) { - return acc + (((value - mean) * (value - mean))/ srcLen); - }; - - float acc = std::accumulate(ptrSrc, ptrSrc + srcLen, 0.0, - VarianceFunction); - return sqrtf(acc); -} - diff --git a/samples/SpeechRecognition/src/Preprocess.cpp b/samples/SpeechRecognition/src/Preprocess.cpp deleted file mode 100644 index 86279619d7..0000000000 --- a/samples/SpeechRecognition/src/Preprocess.cpp +++ /dev/null @@ -1,192 +0,0 @@ -// -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include <algorithm> -#include <numeric> -#include <math.h> -#include <string.h> - -#include "MathUtils.hpp" -#include "Preprocess.hpp" - -Preprocess::Preprocess( - const uint32_t windowLen, - const uint32_t windowStride, - const MFCC mfccInst): - _m_mfcc(mfccInst), - _m_mfccBuf(mfccInst._m_params.m_numMfccFeatures, mfccInst._m_params.m_numMfccVectors), - _m_delta1Buf(mfccInst._m_params.m_numMfccFeatures, mfccInst._m_params.m_numMfccVectors), - _m_delta2Buf(mfccInst._m_params.m_numMfccFeatures, mfccInst._m_params.m_numMfccVectors), - _m_windowLen(windowLen), - _m_windowStride(windowStride) -{ - if (mfccInst._m_params.m_numMfccFeatures > 0 && windowLen > 0) - { - this->_m_mfcc.Init(); - } -} - -Preprocess::~Preprocess() -{ -} - -bool Preprocess::Invoke( const float* audioData, const uint32_t audioDataLen, std::vector<int8_t>& output, - int quantOffset, float quantScale) -{ - this->_m_window = SlidingWindow<const float>( - audioData, audioDataLen, - this->_m_windowLen, this->_m_windowStride); - - uint32_t mfccBufIdx = 0; - - // Init buffers with 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 window */ - while (this->_m_window.HasNext()) - { - const float* mfccWindow = this->_m_window.Next(); - auto mfccAudioData = std::vector<float>( - mfccWindow, - mfccWindow + this->_m_windowLen); - - 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 repeating last feature vector */ - while (mfccBufIdx != this->_m_mfcc._m_params.m_numMfccVectors) - { - memcpy(&this->_m_mfccBuf(0, mfccBufIdx), - &this->_m_mfccBuf(0, mfccBufIdx-1), sizeof(float)*this->_m_mfcc._m_params.m_numMfccFeatures); - ++mfccBufIdx; - } - - /* Compute first and second order deltas from MFCCs */ - this->_ComputeDeltas(this->_m_mfccBuf, - this->_m_delta1Buf, - this->_m_delta2Buf); - - /* Normalise */ - this->_Normalise(); - - return this->_Quantise<int8_t>(output.data(), quantOffset, quantScale); -} - -bool Preprocess::_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; -} - -float Preprocess::_GetMean(Array2d<float>& vec) -{ - return MathUtils::MeanF32(vec.begin(), vec.totalSize()); -} - -float Preprocess::_GetStdDev(Array2d<float>& vec, const float mean) -{ - return MathUtils::StdDevF32(vec.begin(), vec.totalSize(), mean); -} - -void Preprocess::_NormaliseVec(Array2d<float>& vec) -{ - auto mean = Preprocess::_GetMean(vec); - auto stddev = Preprocess::_GetStdDev(vec, mean); - - 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 Preprocess::_Normalise() -{ - Preprocess::_NormaliseVec(this->_m_mfccBuf); - Preprocess::_NormaliseVec(this->_m_delta1Buf); - Preprocess::_NormaliseVec(this->_m_delta2Buf); -} - -float Preprocess::_GetQuantElem( - const float elem, - const float quantScale, - const int quantOffset, - const float minVal, - const float maxVal) -{ - float val = std::round((elem/quantScale) + quantOffset); - float maxim = std::max<float>(val, minVal); - float returnVal = std::min<float>(std::max<float>(val, minVal), maxVal); - return returnVal; -}
\ No newline at end of file diff --git a/samples/SpeechRecognition/src/SpeechRecognitionPipeline.cpp b/samples/SpeechRecognition/src/SpeechRecognitionPipeline.cpp index 1b822d6a88..8b7dd11cb4 100644 --- a/samples/SpeechRecognition/src/SpeechRecognitionPipeline.cpp +++ b/samples/SpeechRecognition/src/SpeechRecognitionPipeline.cpp @@ -6,21 +6,86 @@ #include "SpeechRecognitionPipeline.hpp" #include "ArmnnNetworkExecutor.hpp" -namespace asr +namespace asr { + ASRPipeline::ASRPipeline(std::unique_ptr<common::ArmnnNetworkExecutor<int8_t>> executor, - std::unique_ptr<Decoder> decoder - ) : + std::unique_ptr<Decoder> decoder, std::unique_ptr<Wav2LetterPreprocessor> preProcessor) : m_executor(std::move(executor)), - m_decoder(std::move(decoder)){} + m_decoder(std::move(decoder)), m_preProcessor(std::move(preProcessor)) {} -IPipelinePtr CreatePipeline(common::PipelineOptions& config, std::map<int, std::string>& labels) +int ASRPipeline::getInputSamplesSize() { - auto executor = std::make_unique<common::ArmnnNetworkExecutor<int8_t>>(config.m_ModelFilePath, config.m_backends); + return this->m_preProcessor->m_windowLen + + ((this->m_preProcessor->m_mfcc->m_params.m_numMfccVectors - 1) * this->m_preProcessor->m_windowStride); +} + +int ASRPipeline::getSlidingWindowOffset() +{ + // Hardcoded for now until refactor + return ASRPipeline::SLIDING_WINDOW_OFFSET; +} + +std::vector<int8_t> ASRPipeline::PreProcessing(std::vector<float>& audio) +{ + int audioDataToPreProcess = m_preProcessor->m_windowLen + + ((m_preProcessor->m_mfcc->m_params.m_numMfccVectors - 1) * + m_preProcessor->m_windowStride); + int outputBufferSize = m_preProcessor->m_mfcc->m_params.m_numMfccVectors + * m_preProcessor->m_mfcc->m_params.m_numMfccFeatures * 3; + std::vector<int8_t> outputBuffer(outputBufferSize); + m_preProcessor->Invoke(audio.data(), audioDataToPreProcess, outputBuffer, m_executor->GetQuantizationOffset(), + m_executor->GetQuantizationScale()); + return outputBuffer; +} + +IPipelinePtr CreatePipeline(common::PipelineOptions& config, std::map<int, std::string>& labels) +{ + if (config.m_ModelName == "Wav2Letter") + { + // Wav2Letter ASR SETTINGS + int SAMP_FREQ = 16000; + int FRAME_LEN_MS = 32; + int FRAME_LEN_SAMPLES = SAMP_FREQ * FRAME_LEN_MS * 0.001; + int NUM_MFCC_FEATS = 13; + int MFCC_WINDOW_LEN = 512; + int MFCC_WINDOW_STRIDE = 160; + const int NUM_MFCC_VECTORS = 296; + int SAMPLES_PER_INFERENCE = MFCC_WINDOW_LEN + ((NUM_MFCC_VECTORS - 1) * MFCC_WINDOW_STRIDE); + int MEL_LO_FREQ = 0; + int MEL_HI_FREQ = 8000; + int NUM_FBANK_BIN = 128; + int INPUT_WINDOW_LEFT_CONTEXT = 98; + int INPUT_WINDOW_RIGHT_CONTEXT = 98; + int INPUT_WINDOW_INNER_CONTEXT = NUM_MFCC_VECTORS - + (INPUT_WINDOW_LEFT_CONTEXT + INPUT_WINDOW_RIGHT_CONTEXT); + int SLIDING_WINDOW_OFFSET = INPUT_WINDOW_INNER_CONTEXT * MFCC_WINDOW_STRIDE; + + + MfccParams mfccParams(SAMP_FREQ, NUM_FBANK_BIN, + MEL_LO_FREQ, MEL_HI_FREQ, NUM_MFCC_FEATS, FRAME_LEN_SAMPLES, false, NUM_MFCC_VECTORS); + + std::unique_ptr<Wav2LetterMFCC> mfccInst = std::make_unique<Wav2LetterMFCC>(mfccParams); + + auto executor = std::make_unique<common::ArmnnNetworkExecutor<int8_t>>(config.m_ModelFilePath, + config.m_backends); + + auto decoder = std::make_unique<asr::Decoder>(labels); + + auto preprocessor = std::make_unique<Wav2LetterPreprocessor>(MFCC_WINDOW_LEN, MFCC_WINDOW_STRIDE, + std::move(mfccInst)); + + auto ptr = std::make_unique<asr::ASRPipeline>( + std::move(executor), std::move(decoder), std::move(preprocessor)); - auto decoder = std::make_unique<asr::Decoder>(labels); + ptr->SLIDING_WINDOW_OFFSET = SLIDING_WINDOW_OFFSET; - return std::make_unique<asr::ASRPipeline>(std::move(executor), std::move(decoder)); + return ptr; + } + else + { + throw std::invalid_argument("Unknown Model name: " + config.m_ModelName + " ."); + } } }// namespace asr
\ No newline at end of file diff --git a/samples/SpeechRecognition/src/Wav2LetterMFCC.cpp b/samples/SpeechRecognition/src/Wav2LetterMFCC.cpp new file mode 100644 index 0000000000..959bd9022e --- /dev/null +++ b/samples/SpeechRecognition/src/Wav2LetterMFCC.cpp @@ -0,0 +1,126 @@ +// +// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// +#include "Wav2LetterMFCC.hpp" +#include "MathUtils.hpp" + +#include <cfloat> + +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("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. + 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 * sqrtf(1.0f / + static_cast<float>(4 * inputLength)); + const float normalizer = 2 * sqrtf(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 * cosf((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))); +} diff --git a/samples/SpeechRecognition/src/Wav2LetterPreprocessor.cpp b/samples/SpeechRecognition/src/Wav2LetterPreprocessor.cpp new file mode 100644 index 0000000000..9329d5e4d5 --- /dev/null +++ b/samples/SpeechRecognition/src/Wav2LetterPreprocessor.cpp @@ -0,0 +1,187 @@ +// +// Copyright © 2021 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// +#include "MathUtils.hpp" +#include <cstring> +#include <cmath> +#include <numeric> +#include <algorithm> +#include <memory> +#include "Wav2LetterPreprocessor.hpp" +#include "Wav2LetterMFCC.hpp" + +float Wav2LetterPreprocessor::GetMean(Array2d<float>& vec) +{ + return MathUtils::MeanF32(vec.begin(), vec.totalSize()); +} + +float Wav2LetterPreprocessor::GetStdDev(Array2d<float>& vec, const float mean) +{ + return MathUtils::StdDevF32(vec.begin(), vec.totalSize(), mean); +} + +void Wav2LetterPreprocessor::NormaliseVec(Array2d<float>& vec) +{ + auto mean = Wav2LetterPreprocessor::GetMean(vec); + auto stddev = Wav2LetterPreprocessor::GetStdDev(vec, mean); + + 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 Wav2LetterPreprocessor::Normalise() +{ + Wav2LetterPreprocessor::NormaliseVec(this->m_mfccBuf); + Wav2LetterPreprocessor::NormaliseVec(this->m_delta1Buf); + Wav2LetterPreprocessor::NormaliseVec(this->m_delta2Buf); +} + +float Wav2LetterPreprocessor::GetQuantElem( + const float elem, + const float quantScale, + const int quantOffset, + const float minVal, + const float maxVal) +{ + float val = std::round((elem/quantScale) + quantOffset); + float returnVal = std::min<float>(std::max<float>(val, minVal), maxVal); + return returnVal; +} + +bool Wav2LetterPreprocessor::Invoke(const float* audioData, const uint32_t audioDataLen, std::vector<int8_t>& output, + int quantOffset, float quantScale) +{ + this->m_window = SlidingWindow<const float>( + audioData, audioDataLen, + this->m_windowLen, this->m_windowStride); + + uint32_t mfccBufIdx = 0; + + // Init buffers with 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 window + while (this->m_window.HasNext()) + { + const float* mfccWindow = this->m_window.Next(); + auto mfccAudioData = std::vector<float>( + mfccWindow, + mfccWindow + this->m_windowLen); + + 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 repeating last feature vector + while (mfccBufIdx != this->m_mfcc->m_params.m_numMfccVectors) + { + memcpy(&this->m_mfccBuf(0, mfccBufIdx), + &this->m_mfccBuf(0, mfccBufIdx - 1), sizeof(float) * this->m_mfcc->m_params.m_numMfccFeatures); + ++mfccBufIdx; + } + + // Compute first and second order deltas from MFCCs + Wav2LetterPreprocessor::ComputeDeltas(this->m_mfccBuf, + this->m_delta1Buf, + this->m_delta2Buf); + + // Normalise + this->Normalise(); + + return this->Quantise<int8_t>(output.data(), quantOffset, quantScale); +} + +bool Wav2LetterPreprocessor::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; +} + +Wav2LetterPreprocessor::Wav2LetterPreprocessor(const uint32_t windowLen, + const uint32_t windowStride, + std::unique_ptr<Wav2LetterMFCC> mfccInst): + m_mfcc(std::move(mfccInst)), + m_mfccBuf(m_mfcc->m_params.m_numMfccFeatures, m_mfcc->m_params.m_numMfccVectors), + m_delta1Buf(m_mfcc->m_params.m_numMfccFeatures, m_mfcc->m_params.m_numMfccVectors), + m_delta2Buf(m_mfcc->m_params.m_numMfccFeatures, m_mfcc->m_params.m_numMfccVectors), + m_windowLen(windowLen), + m_windowStride(windowStride) +{ + if (m_mfcc->m_params.m_numMfccFeatures > 0 && windowLen > 0) + { + this->m_mfcc->Init(); + } + std::fill(m_mfccBuf.begin(), m_mfccBuf.end(), 0.f); +}
\ No newline at end of file diff --git a/samples/SpeechRecognition/test/AudioCaptureTest.cpp b/samples/SpeechRecognition/test/AudioCaptureTest.cpp deleted file mode 100644 index 94b4e7cb7a..0000000000 --- a/samples/SpeechRecognition/test/AudioCaptureTest.cpp +++ /dev/null @@ -1,61 +0,0 @@ -// -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#define CATCH_CONFIG_MAIN -#include <catch.hpp> -#include <limits> - -#include "AudioCapture.hpp" - -TEST_CASE("Test capture of audio file") -{ - std::string testResources = TEST_RESOURCE_DIR; - REQUIRE(testResources != ""); - std::string file = testResources + "/" + "myVoiceIsMyPassportVerifyMe04.wav"; - asr::AudioCapture capture; - std::vector<float> audioData = capture.LoadAudioFile(file); - capture.InitSlidingWindow(audioData.data(), audioData.size(), 47712, 16000); - - std::vector<float> firstAudioBlock = capture.Next(); - float actual1 = firstAudioBlock.at(0); - float actual2 = firstAudioBlock.at(47000); - CHECK(std::to_string(actual1) == "0.000352"); - CHECK(std::to_string(actual2) == "-0.056441"); - CHECK(firstAudioBlock.size() == 47712); - - CHECK(capture.HasNext() == true); - - std::vector<float> secondAudioBlock = capture.Next(); - float actual3 = secondAudioBlock.at(0); - float actual4 = secondAudioBlock.at(47000); - CHECK(std::to_string(actual3) == "0.102077"); - CHECK(std::to_string(actual4) == "0.000194"); - CHECK(capture.HasNext() == true); - - std::vector<float> thirdAudioBlock = capture.Next(); - float actual5 = thirdAudioBlock.at(0); - float actual6 = thirdAudioBlock.at(33500); - float actual7 = thirdAudioBlock.at(33600); - CHECK(std::to_string(actual5) == "-0.076416"); - CHECK(std::to_string(actual6) == "-0.000275"); - CHECK(std::to_string(actual7) == "0.000000"); - CHECK(capture.HasNext() == false); -} - -TEST_CASE("Test sliding window of audio capture") -{ - std::string testResources = TEST_RESOURCE_DIR; - REQUIRE(testResources != ""); - std::string file = testResources + "/" + "myVoiceIsMyPassportVerifyMe04.wav"; - asr::AudioCapture capture; - std::vector<float> audioData = capture.LoadAudioFile(file); - capture.InitSlidingWindow(audioData.data(), audioData.size(), 47712, 16000); - capture.Next(); - capture.Next(); - - CHECK(capture.HasNext() == true); - capture.Next(); - CHECK(capture.HasNext() == false); -} diff --git a/samples/SpeechRecognition/test/MFCCTest.cpp b/samples/SpeechRecognition/test/MFCCTest.cpp index 2a552643d5..62a92fd5ba 100644 --- a/samples/SpeechRecognition/test/MFCCTest.cpp +++ b/samples/SpeechRecognition/test/MFCCTest.cpp @@ -6,9 +6,10 @@ #include <catch.hpp> #include <limits> -#include "MFCC.hpp" +#include "Wav2LetterMFCC.hpp" -const std::vector<float> testWav = std::vector<float>{ +const std::vector<float> testWav = std::vector<float> +{ -3.0f, 0.0f, 1.0f, -1.0f, 2.0f, 3.0f, -2.0f, 2.0f, 1.0f, -2.0f, 0.0f, 3.0f, -1.0f, 8.0f, 3.0f, 2.0f, -1.0f, -1.0f, 2.0f, 7.0f, 3.0f, 5.0f, 6.0f, 6.0f, @@ -84,15 +85,16 @@ TEST_CASE("Test MFCC") std::vector<float> fullAudioData; - for (auto f : testWav) - { - fullAudioData.emplace_back( f / (1<<15)); - } - + for (auto f : testWav) + { + fullAudioData.emplace_back( f / (1<<15)); + } - MfccParams mfccParams(sampFreq, 128, 0, 8000, numMfccFeats, frameLenSamples, false, 1); + MfccParams mfccParams(sampFreq, 128, 0, 8000, numMfccFeats, + frameLenSamples, false, 1); - MFCC mfccInst = MFCC(mfccParams); + Wav2LetterMFCC mfccInst = Wav2LetterMFCC(mfccParams); + mfccInst.Init(); auto mfccOutput = mfccInst.MfccCompute(fullAudioData); std::vector<float> expected = { -834.96564f, 21.02699f, 18.62856f, 7.3412f, 18.90791f, -5.36034f, 6.52351f, diff --git a/samples/SpeechRecognition/test/PreprocessTest.cpp b/samples/SpeechRecognition/test/PreprocessTest.cpp index 2b98831fda..f1127470fd 100644 --- a/samples/SpeechRecognition/test/PreprocessTest.cpp +++ b/samples/SpeechRecognition/test/PreprocessTest.cpp @@ -6,8 +6,8 @@ #include <catch.hpp> #include <limits> -#include "Preprocess.hpp" #include "DataStructures.hpp" +#include "Wav2LetterPreprocessor.hpp" void PopulateTestWavVector(std::vector<int16_t>& vec) { @@ -51,9 +51,10 @@ TEST_CASE("Preprocessing calculation INT8") /* Populate with dummy input */ PopulateTestWavVector(testWav1); - MfccParams mfccParams(sampFreq, 128, 0, 8000, numMfccFeats, frameLenSamples, false, numMfccVectors); + MfccParams mfccParams(sampFreq, 128, 0, 8000, numMfccFeats, + frameLenSamples, false, numMfccVectors); - MFCC mfccInst = MFCC(mfccParams); + std::unique_ptr<Wav2LetterMFCC> mfccInst = std::make_unique<Wav2LetterMFCC>(mfccParams); std::vector<float> fullAudioData; @@ -65,7 +66,7 @@ TEST_CASE("Preprocessing calculation INT8") } } - Preprocess prep(frameLenSamples, windowStride, mfccInst); + Wav2LetterPreprocessor prep(frameLenSamples, windowStride, std::move(mfccInst)); std::vector<int8_t> outputBuffer(outputBufferSize); |