Éanna Ó Catháin | c6ab02a | 2021-04-07 14:35:25 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2020 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include <algorithm> |
| 7 | #include <numeric> |
| 8 | #include <math.h> |
| 9 | #include <string.h> |
| 10 | |
| 11 | #include "MathUtils.hpp" |
| 12 | #include "Preprocess.hpp" |
| 13 | |
| 14 | Preprocess::Preprocess( |
| 15 | const uint32_t windowLen, |
| 16 | const uint32_t windowStride, |
| 17 | const MFCC mfccInst): |
| 18 | _m_mfcc(mfccInst), |
| 19 | _m_mfccBuf(mfccInst._m_params.m_numMfccFeatures, mfccInst._m_params.m_numMfccVectors), |
| 20 | _m_delta1Buf(mfccInst._m_params.m_numMfccFeatures, mfccInst._m_params.m_numMfccVectors), |
| 21 | _m_delta2Buf(mfccInst._m_params.m_numMfccFeatures, mfccInst._m_params.m_numMfccVectors), |
| 22 | _m_windowLen(windowLen), |
| 23 | _m_windowStride(windowStride) |
| 24 | { |
| 25 | if (mfccInst._m_params.m_numMfccFeatures > 0 && windowLen > 0) |
| 26 | { |
| 27 | this->_m_mfcc.Init(); |
| 28 | } |
| 29 | } |
| 30 | |
| 31 | Preprocess::~Preprocess() |
| 32 | { |
| 33 | } |
| 34 | |
| 35 | bool Preprocess::Invoke( const float* audioData, const uint32_t audioDataLen, std::vector<int8_t>& output, |
| 36 | int quantOffset, float quantScale) |
| 37 | { |
| 38 | this->_m_window = SlidingWindow<const float>( |
| 39 | audioData, audioDataLen, |
| 40 | this->_m_windowLen, this->_m_windowStride); |
| 41 | |
| 42 | uint32_t mfccBufIdx = 0; |
| 43 | |
| 44 | // Init buffers with 0 |
| 45 | std::fill(_m_mfccBuf.begin(), _m_mfccBuf.end(), 0.f); |
| 46 | std::fill(_m_delta1Buf.begin(), _m_delta1Buf.end(), 0.f); |
| 47 | std::fill(_m_delta2Buf.begin(), _m_delta2Buf.end(), 0.f); |
| 48 | |
| 49 | /* While we can slide over the window */ |
| 50 | while (this->_m_window.HasNext()) |
| 51 | { |
| 52 | const float* mfccWindow = this->_m_window.Next(); |
| 53 | auto mfccAudioData = std::vector<float>( |
| 54 | mfccWindow, |
| 55 | mfccWindow + this->_m_windowLen); |
| 56 | |
| 57 | auto mfcc = this->_m_mfcc.MfccCompute(mfccAudioData); |
| 58 | for (size_t i = 0; i < this->_m_mfccBuf.size(0); ++i) |
| 59 | { |
| 60 | this->_m_mfccBuf(i, mfccBufIdx) = mfcc[i]; |
| 61 | } |
| 62 | ++mfccBufIdx; |
| 63 | } |
| 64 | |
| 65 | /* Pad MFCC if needed by repeating last feature vector */ |
| 66 | while (mfccBufIdx != this->_m_mfcc._m_params.m_numMfccVectors) |
| 67 | { |
| 68 | memcpy(&this->_m_mfccBuf(0, mfccBufIdx), |
| 69 | &this->_m_mfccBuf(0, mfccBufIdx-1), sizeof(float)*this->_m_mfcc._m_params.m_numMfccFeatures); |
| 70 | ++mfccBufIdx; |
| 71 | } |
| 72 | |
| 73 | /* Compute first and second order deltas from MFCCs */ |
| 74 | this->_ComputeDeltas(this->_m_mfccBuf, |
| 75 | this->_m_delta1Buf, |
| 76 | this->_m_delta2Buf); |
| 77 | |
| 78 | /* Normalise */ |
| 79 | this->_Normalise(); |
| 80 | |
| 81 | return this->_Quantise<int8_t>(output.data(), quantOffset, quantScale); |
| 82 | } |
| 83 | |
| 84 | bool Preprocess::_ComputeDeltas(Array2d<float>& mfcc, |
| 85 | Array2d<float>& delta1, |
| 86 | Array2d<float>& delta2) |
| 87 | { |
| 88 | const std::vector <float> delta1Coeffs = |
| 89 | {6.66666667e-02, 5.00000000e-02, 3.33333333e-02, |
| 90 | 1.66666667e-02, -3.46944695e-18, -1.66666667e-02, |
| 91 | -3.33333333e-02, -5.00000000e-02, -6.66666667e-02}; |
| 92 | |
| 93 | const std::vector <float> delta2Coeffs = |
| 94 | {0.06060606, 0.01515152, -0.01731602, |
| 95 | -0.03679654, -0.04329004, -0.03679654, |
| 96 | -0.01731602, 0.01515152, 0.06060606}; |
| 97 | |
| 98 | if (delta1.size(0) == 0 || delta2.size(0) != delta1.size(0) || |
| 99 | mfcc.size(0) == 0 || mfcc.size(1) == 0) |
| 100 | { |
| 101 | return false; |
| 102 | } |
| 103 | |
| 104 | /* Get the middle index; coeff vec len should always be odd */ |
| 105 | const size_t coeffLen = delta1Coeffs.size(); |
| 106 | const size_t fMidIdx = (coeffLen - 1)/2; |
| 107 | const size_t numFeatures = mfcc.size(0); |
| 108 | const size_t numFeatVectors = mfcc.size(1); |
| 109 | |
| 110 | /* iterate through features in MFCC vector*/ |
| 111 | for (size_t i = 0; i < numFeatures; ++i) |
| 112 | { |
| 113 | /* for each feature, iterate through time (t) samples representing feature evolution and |
| 114 | * calculate d/dt and d^2/dt^2, using 1d convolution with differential kernels. |
| 115 | * Convolution padding = valid, result size is `time length - kernel length + 1`. |
| 116 | * The result is padded with 0 from both sides to match the size of initial time samples data. |
| 117 | * |
| 118 | * For the small filter, conv1d implementation as a simple loop is efficient enough. |
| 119 | * Filters of a greater size would need CMSIS-DSP functions to be used, like arm_fir_f32. |
| 120 | */ |
| 121 | |
| 122 | for (size_t j = fMidIdx; j < numFeatVectors - fMidIdx; ++j) |
| 123 | { |
| 124 | float d1 = 0; |
| 125 | float d2 = 0; |
| 126 | const size_t mfccStIdx = j - fMidIdx; |
| 127 | |
| 128 | for (size_t k = 0, m = coeffLen - 1; k < coeffLen; ++k, --m) |
| 129 | { |
| 130 | |
| 131 | d1 += mfcc(i,mfccStIdx + k) * delta1Coeffs[m]; |
| 132 | d2 += mfcc(i,mfccStIdx + k) * delta2Coeffs[m]; |
| 133 | } |
| 134 | |
| 135 | delta1(i,j) = d1; |
| 136 | delta2(i,j) = d2; |
| 137 | } |
| 138 | } |
| 139 | |
| 140 | return true; |
| 141 | } |
| 142 | |
| 143 | float Preprocess::_GetMean(Array2d<float>& vec) |
| 144 | { |
| 145 | return MathUtils::MeanF32(vec.begin(), vec.totalSize()); |
| 146 | } |
| 147 | |
| 148 | float Preprocess::_GetStdDev(Array2d<float>& vec, const float mean) |
| 149 | { |
| 150 | return MathUtils::StdDevF32(vec.begin(), vec.totalSize(), mean); |
| 151 | } |
| 152 | |
| 153 | void Preprocess::_NormaliseVec(Array2d<float>& vec) |
| 154 | { |
| 155 | auto mean = Preprocess::_GetMean(vec); |
| 156 | auto stddev = Preprocess::_GetStdDev(vec, mean); |
| 157 | |
| 158 | if (stddev == 0) |
| 159 | { |
| 160 | std::fill(vec.begin(), vec.end(), 0); |
| 161 | } |
| 162 | else |
| 163 | { |
| 164 | const float stddevInv = 1.f/stddev; |
| 165 | const float normalisedMean = mean/stddev; |
| 166 | |
| 167 | auto NormalisingFunction = [=](float &value) { |
| 168 | value = value * stddevInv - normalisedMean; |
| 169 | }; |
| 170 | std::for_each(vec.begin(), vec.end(), NormalisingFunction); |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | void Preprocess::_Normalise() |
| 175 | { |
| 176 | Preprocess::_NormaliseVec(this->_m_mfccBuf); |
| 177 | Preprocess::_NormaliseVec(this->_m_delta1Buf); |
| 178 | Preprocess::_NormaliseVec(this->_m_delta2Buf); |
| 179 | } |
| 180 | |
| 181 | float Preprocess::_GetQuantElem( |
| 182 | const float elem, |
| 183 | const float quantScale, |
| 184 | const int quantOffset, |
| 185 | const float minVal, |
| 186 | const float maxVal) |
| 187 | { |
| 188 | float val = std::round((elem/quantScale) + quantOffset); |
| 189 | float maxim = std::max<float>(val, minVal); |
| 190 | float returnVal = std::min<float>(std::max<float>(val, minVal), maxVal); |
| 191 | return returnVal; |
| 192 | } |