diff options
Diffstat (limited to 'samples/SpeechRecognition/src')
-rw-r--r-- | samples/SpeechRecognition/src/AudioCapture.cpp | 104 | ||||
-rw-r--r-- | samples/SpeechRecognition/src/Decoder.cpp | 45 | ||||
-rw-r--r-- | samples/SpeechRecognition/src/MFCC.cpp | 397 | ||||
-rw-r--r-- | samples/SpeechRecognition/src/Main.cpp | 137 | ||||
-rw-r--r-- | samples/SpeechRecognition/src/MathUtils.cpp | 112 | ||||
-rw-r--r-- | samples/SpeechRecognition/src/Preprocess.cpp | 192 | ||||
-rw-r--r-- | samples/SpeechRecognition/src/SpeechRecognitionPipeline.cpp | 81 | ||||
-rw-r--r-- | samples/SpeechRecognition/src/Wav2LetterMFCC.cpp | 126 | ||||
-rw-r--r-- | samples/SpeechRecognition/src/Wav2LetterPreprocessor.cpp | 187 |
9 files changed, 464 insertions, 917 deletions
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 |