alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 Arm Limited. All rights reserved. |
| 3 | * SPDX-License-Identifier: Apache-2.0 |
| 4 | * |
| 5 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | * you may not use this file except in compliance with the License. |
| 7 | * You may obtain a copy of the License at |
| 8 | * |
| 9 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 10 | * |
| 11 | * Unless required by applicable law or agreed to in writing, software |
| 12 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 13 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | * See the License for the specific language governing permissions and |
| 15 | * limitations under the License. |
| 16 | */ |
| 17 | #include "Wav2LetterPreprocess.hpp" |
| 18 | |
| 19 | #include "PlatformMath.hpp" |
| 20 | #include "TensorFlowLiteMicro.hpp" |
| 21 | |
| 22 | #include <algorithm> |
| 23 | #include <math.h> |
| 24 | |
| 25 | namespace arm { |
| 26 | namespace app { |
| 27 | namespace audio { |
| 28 | namespace asr { |
| 29 | |
| 30 | Preprocess::Preprocess( |
| 31 | const uint32_t numMfccFeatures, |
| 32 | const uint32_t windowLen, |
| 33 | const uint32_t windowStride, |
| 34 | const uint32_t numMfccVectors): |
| 35 | _m_mfcc(numMfccFeatures, windowLen), |
| 36 | _m_mfccBuf(numMfccFeatures, numMfccVectors), |
| 37 | _m_delta1Buf(numMfccFeatures, numMfccVectors), |
| 38 | _m_delta2Buf(numMfccFeatures, numMfccVectors), |
| 39 | _m_windowLen(windowLen), |
| 40 | _m_windowStride(windowStride), |
| 41 | _m_numMfccFeats(numMfccFeatures), |
| 42 | _m_numFeatVectors(numMfccVectors), |
| 43 | _m_window() |
| 44 | { |
| 45 | if (numMfccFeatures > 0 && windowLen > 0) { |
| 46 | this->_m_mfcc.Init(); |
| 47 | } |
| 48 | } |
| 49 | |
| 50 | bool Preprocess::Invoke( |
| 51 | const int16_t* audioData, |
| 52 | const uint32_t audioDataLen, |
| 53 | TfLiteTensor* tensor) |
| 54 | { |
| 55 | this->_m_window = SlidingWindow<const int16_t>( |
| 56 | audioData, audioDataLen, |
| 57 | this->_m_windowLen, this->_m_windowStride); |
| 58 | |
| 59 | uint32_t mfccBufIdx = 0; |
| 60 | |
| 61 | std::fill(_m_mfccBuf.begin(), _m_mfccBuf.end(), 0.f); |
| 62 | std::fill(_m_delta1Buf.begin(), _m_delta1Buf.end(), 0.f); |
| 63 | std::fill(_m_delta2Buf.begin(), _m_delta2Buf.end(), 0.f); |
| 64 | |
| 65 | /* While we can slide over the window. */ |
| 66 | while (this->_m_window.HasNext()) { |
| 67 | const int16_t* mfccWindow = this->_m_window.Next(); |
| 68 | auto mfccAudioData = std::vector<int16_t>( |
| 69 | mfccWindow, |
| 70 | mfccWindow + this->_m_windowLen); |
| 71 | auto mfcc = this->_m_mfcc.MfccCompute(mfccAudioData); |
| 72 | for (size_t i = 0; i < this->_m_mfccBuf.size(0); ++i) { |
| 73 | this->_m_mfccBuf(i, mfccBufIdx) = mfcc[i]; |
| 74 | } |
| 75 | ++mfccBufIdx; |
| 76 | } |
| 77 | |
| 78 | /* Pad MFCC if needed by adding MFCC for zeros. */ |
| 79 | if (mfccBufIdx != this->_m_numFeatVectors) { |
| 80 | std::vector<int16_t> zerosWindow = std::vector<int16_t>(this->_m_windowLen, 0); |
| 81 | std::vector<float> mfccZeros = this->_m_mfcc.MfccCompute(zerosWindow); |
| 82 | |
| 83 | while (mfccBufIdx != this->_m_numFeatVectors) { |
| 84 | memcpy(&this->_m_mfccBuf(0, mfccBufIdx), |
| 85 | mfccZeros.data(), sizeof(float) * _m_numMfccFeats); |
| 86 | ++mfccBufIdx; |
| 87 | } |
| 88 | } |
| 89 | |
| 90 | /* Compute first and second order deltas from MFCCs. */ |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 91 | this->ComputeDeltas(this->_m_mfccBuf, |
| 92 | this->_m_delta1Buf, |
| 93 | this->_m_delta2Buf); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 94 | |
| 95 | /* Normalise. */ |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 96 | this->Normalise(); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 97 | |
| 98 | /* Quantise. */ |
| 99 | QuantParams quantParams = GetTensorQuantParams(tensor); |
| 100 | |
| 101 | if (0 == quantParams.scale) { |
| 102 | printf_err("Quantisation scale can't be 0\n"); |
| 103 | return false; |
| 104 | } |
| 105 | |
| 106 | switch(tensor->type) { |
| 107 | case kTfLiteUInt8: |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 108 | return this->Quantise<uint8_t>( |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 109 | tflite::GetTensorData<uint8_t>(tensor), tensor->bytes, |
| 110 | quantParams.scale, quantParams.offset); |
| 111 | case kTfLiteInt8: |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 112 | return this->Quantise<int8_t>( |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 113 | tflite::GetTensorData<int8_t>(tensor), tensor->bytes, |
| 114 | quantParams.scale, quantParams.offset); |
| 115 | default: |
| 116 | printf_err("Unsupported tensor type %s\n", |
| 117 | TfLiteTypeGetName(tensor->type)); |
| 118 | } |
| 119 | |
| 120 | return false; |
| 121 | } |
| 122 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 123 | bool Preprocess::ComputeDeltas(Array2d<float>& mfcc, |
| 124 | Array2d<float>& delta1, |
| 125 | Array2d<float>& delta2) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 126 | { |
| 127 | const std::vector <float> delta1Coeffs = |
| 128 | {6.66666667e-02, 5.00000000e-02, 3.33333333e-02, |
| 129 | 1.66666667e-02, -3.46944695e-18, -1.66666667e-02, |
| 130 | -3.33333333e-02, -5.00000000e-02, -6.66666667e-02}; |
| 131 | |
| 132 | const std::vector <float> delta2Coeffs = |
| 133 | {0.06060606, 0.01515152, -0.01731602, |
| 134 | -0.03679654, -0.04329004, -0.03679654, |
| 135 | -0.01731602, 0.01515152, 0.06060606}; |
| 136 | |
| 137 | if (delta1.size(0) == 0 || delta2.size(0) != delta1.size(0) || |
| 138 | mfcc.size(0) == 0 || mfcc.size(1) == 0) { |
| 139 | return false; |
| 140 | } |
| 141 | |
| 142 | /* Get the middle index; coeff vec len should always be odd. */ |
| 143 | const size_t coeffLen = delta1Coeffs.size(); |
| 144 | const size_t fMidIdx = (coeffLen - 1)/2; |
| 145 | const size_t numFeatures = mfcc.size(0); |
| 146 | const size_t numFeatVectors = mfcc.size(1); |
| 147 | |
| 148 | /* Iterate through features in MFCC vector. */ |
| 149 | for (size_t i = 0; i < numFeatures; ++i) { |
| 150 | /* For each feature, iterate through time (t) samples representing feature evolution and |
| 151 | * calculate d/dt and d^2/dt^2, using 1d convolution with differential kernels. |
| 152 | * Convolution padding = valid, result size is `time length - kernel length + 1`. |
| 153 | * The result is padded with 0 from both sides to match the size of initial time samples data. |
| 154 | * |
| 155 | * For the small filter, conv1d implementation as a simple loop is efficient enough. |
| 156 | * Filters of a greater size would need CMSIS-DSP functions to be used, like arm_fir_f32. |
| 157 | */ |
| 158 | |
| 159 | for (size_t j = fMidIdx; j < numFeatVectors - fMidIdx; ++j) { |
| 160 | float d1 = 0; |
| 161 | float d2 = 0; |
| 162 | const size_t mfccStIdx = j - fMidIdx; |
| 163 | |
| 164 | for (size_t k = 0, m = coeffLen - 1; k < coeffLen; ++k, --m) { |
| 165 | |
| 166 | d1 += mfcc(i,mfccStIdx + k) * delta1Coeffs[m]; |
| 167 | d2 += mfcc(i,mfccStIdx + k) * delta2Coeffs[m]; |
| 168 | } |
| 169 | |
| 170 | delta1(i,j) = d1; |
| 171 | delta2(i,j) = d2; |
| 172 | } |
| 173 | } |
| 174 | |
| 175 | return true; |
| 176 | } |
| 177 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 178 | float Preprocess::GetMean(Array2d<float>& vec) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 179 | { |
| 180 | return math::MathUtils::MeanF32(vec.begin(), vec.totalSize()); |
| 181 | } |
| 182 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 183 | float Preprocess::GetStdDev(Array2d<float>& vec, const float mean) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 184 | { |
| 185 | return math::MathUtils::StdDevF32(vec.begin(), vec.totalSize(), mean); |
| 186 | } |
| 187 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 188 | void Preprocess::NormaliseVec(Array2d<float>& vec) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 189 | { |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 190 | auto mean = Preprocess::GetMean(vec); |
| 191 | auto stddev = Preprocess::GetStdDev(vec, mean); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 192 | |
| 193 | debug("Mean: %f, Stddev: %f\n", mean, stddev); |
| 194 | if (stddev == 0) { |
| 195 | std::fill(vec.begin(), vec.end(), 0); |
| 196 | } else { |
| 197 | const float stddevInv = 1.f/stddev; |
| 198 | const float normalisedMean = mean/stddev; |
| 199 | |
| 200 | auto NormalisingFunction = [=](float& value) { |
| 201 | value = value * stddevInv - normalisedMean; |
| 202 | }; |
| 203 | std::for_each(vec.begin(), vec.end(), NormalisingFunction); |
| 204 | } |
| 205 | } |
| 206 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 207 | void Preprocess::Normalise() |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 208 | { |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 209 | Preprocess::NormaliseVec(this->_m_mfccBuf); |
| 210 | Preprocess::NormaliseVec(this->_m_delta1Buf); |
| 211 | Preprocess::NormaliseVec(this->_m_delta2Buf); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 212 | } |
| 213 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame^] | 214 | float Preprocess::GetQuantElem( |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 215 | const float elem, |
| 216 | const float quantScale, |
| 217 | const int quantOffset, |
| 218 | const float minVal, |
| 219 | const float maxVal) |
| 220 | { |
| 221 | float val = std::round((elem/quantScale) + quantOffset); |
| 222 | return std::min<float>(std::max<float>(val, minVal), maxVal); |
| 223 | } |
| 224 | |
| 225 | } /* namespace asr */ |
| 226 | } /* namespace audio */ |
| 227 | } /* namespace app */ |
| 228 | } /* namespace arm */ |