diff options
author | Kshitij Sisodia <kshitij.sisodia@arm.com> | 2022-05-06 09:13:03 +0100 |
---|---|---|
committer | Kshitij Sisodia <kshitij.sisodia@arm.com> | 2022-05-06 17:11:41 +0100 |
commit | aa4bcb14d0cbee910331545dd2fc086b58c37170 (patch) | |
tree | e67a43a43f61c6f8b6aad19018b0827baf7e31a6 /source/use_case/ad/src | |
parent | fcca863bafd5f33522bc14c23dde4540e264ec94 (diff) | |
download | ml-embedded-evaluation-kit-aa4bcb14d0cbee910331545dd2fc086b58c37170.tar.gz |
MLECO-3183: Refactoring application sources
Platform agnostic application sources are moved into application
api module with their own independent CMake projects.
Changes for MLECO-3080 also included - they create CMake projects
individial API's (again, platform agnostic) that dependent on the
common logic. The API for KWS_API "joint" API has been removed and
now the use case relies on individual KWS, and ASR API libraries.
Change-Id: I1f7748dc767abb3904634a04e0991b74ac7b756d
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>
Diffstat (limited to 'source/use_case/ad/src')
-rw-r--r-- | source/use_case/ad/src/AdMelSpectrogram.cc | 93 | ||||
-rw-r--r-- | source/use_case/ad/src/AdModel.cc | 54 | ||||
-rw-r--r-- | source/use_case/ad/src/AdProcessing.cc | 208 | ||||
-rw-r--r-- | source/use_case/ad/src/MainLoop.cc | 25 | ||||
-rw-r--r-- | source/use_case/ad/src/MelSpectrogram.cc | 316 |
5 files changed, 23 insertions, 673 deletions
diff --git a/source/use_case/ad/src/AdMelSpectrogram.cc b/source/use_case/ad/src/AdMelSpectrogram.cc deleted file mode 100644 index 14b9323..0000000 --- a/source/use_case/ad/src/AdMelSpectrogram.cc +++ /dev/null @@ -1,93 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "AdMelSpectrogram.hpp" -#include "PlatformMath.hpp" -#include "log_macros.h" - -#include <cfloat> - -namespace arm { -namespace app { -namespace audio { - - bool AdMelSpectrogram::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_err("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(); - float melEnergy = FLT_MIN; /* Avoid log of zero at later stages. */ - const uint32_t firstIndex = filterBankFilterFirst[bin]; - const uint32_t lastIndex = std::min<int32_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 AdMelSpectrogram::ConvertToLogarithmicScale( - std::vector<float>& melEnergies) - { - /* 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 */ - math::MathUtils::VecLogarithmF32(melEnergies, vecLogEnergies); - - /* Scale the log values. */ - for (auto iterM = melEnergies.begin(), iterL = vecLogEnergies.begin(); - iterM != melEnergies.end() && iterL != vecLogEnergies.end(); ++iterM, ++iterL) { - - *iterM = *iterL * multiplier; - } - } - - float AdMelSpectrogram::GetMelFilterBankNormaliser( - const float& leftMel, - const float& rightMel, - const bool useHTKMethod) - { - /* Slaney normalization for mel weights. */ - return (2.0f / (AdMelSpectrogram::InverseMelScale(rightMel, useHTKMethod) - - AdMelSpectrogram::InverseMelScale(leftMel, useHTKMethod))); - } - -} /* namespace audio */ -} /* namespace app */ -} /* namespace arm */ diff --git a/source/use_case/ad/src/AdModel.cc b/source/use_case/ad/src/AdModel.cc deleted file mode 100644 index a2ef260..0000000 --- a/source/use_case/ad/src/AdModel.cc +++ /dev/null @@ -1,54 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "AdModel.hpp" -#include "log_macros.h" - -const tflite::MicroOpResolver& arm::app::AdModel::GetOpResolver() -{ - return this->m_opResolver; -} - -bool arm::app::AdModel::EnlistOperations() -{ - this->m_opResolver.AddAveragePool2D(); - this->m_opResolver.AddConv2D(); - this->m_opResolver.AddDepthwiseConv2D(); - this->m_opResolver.AddRelu6(); - this->m_opResolver.AddReshape(); - -#if defined(ARM_NPU) - if (kTfLiteOk == this->m_opResolver.AddEthosU()) { - info("Added %s support to op resolver\n", - tflite::GetString_ETHOSU()); - } else { - printf_err("Failed to add Arm NPU support to op resolver."); - return false; - } -#endif /* ARM_NPU */ - return true; -} - -extern uint8_t* GetModelPointer(); -const uint8_t* arm::app::AdModel::ModelPointer() -{ - return GetModelPointer(); -} -extern size_t GetModelLen(); -size_t arm::app::AdModel::ModelSize() -{ - return GetModelLen(); -} diff --git a/source/use_case/ad/src/AdProcessing.cc b/source/use_case/ad/src/AdProcessing.cc deleted file mode 100644 index a33131c..0000000 --- a/source/use_case/ad/src/AdProcessing.cc +++ /dev/null @@ -1,208 +0,0 @@ -/* - * Copyright (c) 2022 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "AdProcessing.hpp" - -#include "AdModel.hpp" - -namespace arm { -namespace app { - -AdPreProcess::AdPreProcess(TfLiteTensor* inputTensor, - uint32_t melSpectrogramFrameLen, - uint32_t melSpectrogramFrameStride, - float adModelTrainingMean): - m_validInstance{false}, - m_melSpectrogramFrameLen{melSpectrogramFrameLen}, - m_melSpectrogramFrameStride{melSpectrogramFrameStride}, - /**< Model is trained on features downsampled 2x */ - m_inputResizeScale{2}, - /**< We are choosing to move by 20 frames across the audio for each inference. */ - m_numMelSpecVectorsInAudioStride{20}, - m_audioDataStride{m_numMelSpecVectorsInAudioStride * melSpectrogramFrameStride}, - m_melSpec{melSpectrogramFrameLen} -{ - if (!inputTensor) { - printf_err("Invalid input tensor provided to pre-process\n"); - return; - } - - TfLiteIntArray* inputShape = inputTensor->dims; - - if (!inputShape) { - printf_err("Invalid input tensor dims\n"); - return; - } - - const uint32_t kNumRows = inputShape->data[AdModel::ms_inputRowsIdx]; - const uint32_t kNumCols = inputShape->data[AdModel::ms_inputColsIdx]; - - /* Deduce the data length required for 1 inference from the network parameters. */ - this->m_audioDataWindowSize = (((this->m_inputResizeScale * kNumCols) - 1) * - melSpectrogramFrameStride) + - melSpectrogramFrameLen; - this->m_numReusedFeatureVectors = kNumRows - - (this->m_numMelSpecVectorsInAudioStride / - this->m_inputResizeScale); - this->m_melSpec.Init(); - - /* Creating a Mel Spectrogram sliding window for the data required for 1 inference. - * "resizing" done here by multiplying stride by resize scale. */ - this->m_melWindowSlider = audio::SlidingWindow<const int16_t>( - nullptr, /* to be populated later. */ - this->m_audioDataWindowSize, - melSpectrogramFrameLen, - melSpectrogramFrameStride * this->m_inputResizeScale); - - /* Construct feature calculation function. */ - this->m_featureCalc = GetFeatureCalculator(this->m_melSpec, inputTensor, - this->m_numReusedFeatureVectors, - adModelTrainingMean); - this->m_validInstance = true; -} - -bool AdPreProcess::DoPreProcess(const void* input, size_t inputSize) -{ - /* Check that we have a valid instance. */ - if (!this->m_validInstance) { - printf_err("Invalid pre-processor instance\n"); - return false; - } - - /* We expect that we can traverse the size with which the MEL spectrogram - * sliding window was initialised with. */ - if (!input || inputSize < this->m_audioDataWindowSize) { - printf_err("Invalid input provided for pre-processing\n"); - return false; - } - - /* We moved to the next window - set the features sliding to the new address. */ - this->m_melWindowSlider.Reset(static_cast<const int16_t*>(input)); - - /* The first window does not have cache ready. */ - const bool useCache = this->m_audioWindowIndex > 0 && this->m_numReusedFeatureVectors > 0; - - /* Start calculating features inside one audio sliding window. */ - while (this->m_melWindowSlider.HasNext()) { - const int16_t* melSpecWindow = this->m_melWindowSlider.Next(); - std::vector<int16_t> melSpecAudioData = std::vector<int16_t>( - melSpecWindow, - melSpecWindow + this->m_melSpectrogramFrameLen); - - /* Compute features for this window and write them to input tensor. */ - this->m_featureCalc(melSpecAudioData, - this->m_melWindowSlider.Index(), - useCache, - this->m_numMelSpecVectorsInAudioStride, - this->m_inputResizeScale); - } - - return true; -} - -uint32_t AdPreProcess::GetAudioWindowSize() -{ - return this->m_audioDataWindowSize; -} - -uint32_t AdPreProcess::GetAudioDataStride() -{ - return this->m_audioDataStride; -} - -void AdPreProcess::SetAudioWindowIndex(uint32_t idx) -{ - this->m_audioWindowIndex = idx; -} - -AdPostProcess::AdPostProcess(TfLiteTensor* outputTensor) : - m_outputTensor {outputTensor} -{} - -bool AdPostProcess::DoPostProcess() -{ - switch (this->m_outputTensor->type) { - case kTfLiteInt8: - this->Dequantize<int8_t>(); - break; - default: - printf_err("Unsupported tensor type"); - return false; - } - - math::MathUtils::SoftmaxF32(this->m_dequantizedOutputVec); - return true; -} - -float AdPostProcess::GetOutputValue(uint32_t index) -{ - if (index < this->m_dequantizedOutputVec.size()) { - return this->m_dequantizedOutputVec[index]; - } - printf_err("Invalid index for output\n"); - return 0.0; -} - -std::function<void (std::vector<int16_t>&, int, bool, size_t, size_t)> -GetFeatureCalculator(audio::AdMelSpectrogram& melSpec, - TfLiteTensor* inputTensor, - size_t cacheSize, - float trainingMean) -{ - std::function<void (std::vector<int16_t>&, size_t, bool, size_t, size_t)> melSpecFeatureCalc; - - TfLiteQuantization quant = inputTensor->quantization; - - if (kTfLiteAffineQuantization == quant.type) { - - auto* quantParams = static_cast<TfLiteAffineQuantization*>(quant.params); - const float quantScale = quantParams->scale->data[0]; - const int quantOffset = quantParams->zero_point->data[0]; - - switch (inputTensor->type) { - case kTfLiteInt8: { - melSpecFeatureCalc = FeatureCalc<int8_t>( - inputTensor, - cacheSize, - [=, &melSpec](std::vector<int16_t>& audioDataWindow) { - return melSpec.MelSpecComputeQuant<int8_t>( - audioDataWindow, - quantScale, - quantOffset, - trainingMean); - } - ); - break; - } - default: - printf_err("Tensor type %s not supported\n", TfLiteTypeGetName(inputTensor->type)); - } - } else { - melSpecFeatureCalc = FeatureCalc<float>( - inputTensor, - cacheSize, - [=, &melSpec]( - std::vector<int16_t>& audioDataWindow) { - return melSpec.ComputeMelSpec( - audioDataWindow, - trainingMean); - }); - } - return melSpecFeatureCalc; -} - -} /* namespace app */ -} /* namespace arm */ diff --git a/source/use_case/ad/src/MainLoop.cc b/source/use_case/ad/src/MainLoop.cc index 140359b..e9f7b4e 100644 --- a/source/use_case/ad/src/MainLoop.cc +++ b/source/use_case/ad/src/MainLoop.cc @@ -18,7 +18,17 @@ #include "AdModel.hpp" /* Model class for running inference */ #include "UseCaseCommonUtils.hpp" /* Utils functions */ #include "UseCaseHandler.hpp" /* Handlers for different user options */ -#include "log_macros.h" +#include "log_macros.h" /* Logging functions */ +#include "BufAttributes.hpp" /* Buffer attributes to be applied */ + +namespace arm { + namespace app { + static uint8_t tensorArena[ACTIVATION_BUF_SZ] ACTIVATION_BUF_ATTRIBUTE; + } /* namespace app */ +} /* namespace arm */ + +extern uint8_t* GetModelPointer(); +extern size_t GetModelLen(); enum opcodes { @@ -49,12 +59,23 @@ void main_loop() arm::app::AdModel model; /* Model wrapper object. */ /* Load the model. */ - if (!model.Init()) + if (!model.Init(arm::app::tensorArena, + sizeof(arm::app::tensorArena), + GetModelPointer(), + GetModelLen())) { printf_err("failed to initialise model\n"); return; } +#if !defined(ARM_NPU) + /* If it is not a NPU build check if the model contains a NPU operator */ + if (model.ContainsEthosUOperator()) { + printf_err("No driver support for Ethos-U operator found in the model.\n"); + return; + } +#endif /* ARM_NPU */ + /* Instantiate application context. */ arm::app::ApplicationContext caseContext; diff --git a/source/use_case/ad/src/MelSpectrogram.cc b/source/use_case/ad/src/MelSpectrogram.cc deleted file mode 100644 index ff0c536..0000000 --- a/source/use_case/ad/src/MelSpectrogram.cc +++ /dev/null @@ -1,316 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "MelSpectrogram.hpp" - -#include "PlatformMath.hpp" -#include "log_macros.h" - -#include <cfloat> -#include <cinttypes> - -namespace arm { -namespace app { -namespace audio { - - MelSpecParams::MelSpecParams( - const float samplingFreq, - const uint32_t numFbankBins, - const float melLoFreq, - const float melHiFreq, - const uint32_t frameLen, - const bool useHtkMethod): - m_samplingFreq(samplingFreq), - m_numFbankBins(numFbankBins), - m_melLoFreq(melLoFreq), - m_melHiFreq(melHiFreq), - m_frameLen(frameLen), - - /* Smallest power of 2 >= frame length. */ - m_frameLenPadded(pow(2, ceil((log(frameLen)/log(2))))), - m_useHtkMethod(useHtkMethod) - {} - - std::string MelSpecParams::Str() const - { - char strC[1024]; - snprintf(strC, sizeof(strC) - 1, "\n \ - \n\t Sampling frequency: %f\ - \n\t Number of filter banks: %" PRIu32 "\ - \n\t Mel frequency limit (low): %f\ - \n\t Mel frequency limit (high): %f\ - \n\t Frame length: %" PRIu32 "\ - \n\t Padded frame length: %" PRIu32 "\ - \n\t Using HTK for Mel scale: %s\n", - this->m_samplingFreq, this->m_numFbankBins, this->m_melLoFreq, - this->m_melHiFreq, this->m_frameLen, - this->m_frameLenPadded, this->m_useHtkMethod ? "yes" : "no"); - return std::string{strC}; - } - - MelSpectrogram::MelSpectrogram(const MelSpecParams& 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 auto multiplier = static_cast<float>(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 * - math::MathUtils::CosineF32(static_cast<float>(i) * multiplier))); - } - - math::MathUtils::FftInitF32(this->m_params.m_frameLenPadded, this->m_fftInstance); - debug("Instantiated Mel Spectrogram object: %s\n", this->m_params.Str().c_str()); - } - - void MelSpectrogram::Init() - { - this->InitMelFilterBank(); - } - - float MelSpectrogram::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 / ms_freqStep; - - if (freq >= ms_minLogHz) { - mel = ms_minLogMel + logf(freq / ms_minLogHz) / ms_logStep; - } - return mel; - } - } - - float MelSpectrogram::InverseMelScale(const float melFreq, const bool useHTKMethod) - { - if (useHTKMethod) { - return 700.0f * (expf (melFreq / 1127.0f) - 1.0f); - } else { - /* Slaney formula for inverse mel scale. */ - float freq = ms_freqStep * melFreq; - - if (melFreq >= ms_minLogMel) { - freq = ms_minLogHz * expf(ms_logStep * (melFreq - ms_minLogMel)); - } - return freq; - } - } - - bool MelSpectrogram::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_err("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(); - float melEnergy = FLT_MIN; /* Avoid log of zero at later stages */ - const uint32_t firstIndex = filterBankFilterFirst[bin]; - const uint32_t lastIndex = std::min<int32_t>(filterBankFilterLast[bin], fftVec.size() - 1); - - for (uint32_t i = firstIndex; i <= lastIndex && filterBankIter != end; ++i) { - float energyRep = math::MathUtils::SqrtF32(fftVec[i]); - melEnergy += (*filterBankIter++ * energyRep); - } - - melEnergies[bin] = melEnergy; - } - - return true; - } - - void MelSpectrogram::ConvertToLogarithmicScale(std::vector<float>& melEnergies) - { - for (float& melEnergy : melEnergies) { - melEnergy = logf(melEnergy); - } - } - - void MelSpectrogram::ConvertToPowerSpectrum() - { - const uint32_t halfDim = this->m_buffer.size() / 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]; - - math::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; - } - - float MelSpectrogram::GetMelFilterBankNormaliser( - const float& leftMel, - const float& rightMel, - const bool useHTKMethod) - { - UNUSED(leftMel); - UNUSED(rightMel); - UNUSED(useHTKMethod); - - /* By default, no normalisation => return 1 */ - return 1.f; - } - - void MelSpectrogram::InitMelFilterBank() - { - if (!this->IsMelFilterBankInited()) { - this->m_melFilterBank = this->CreateMelFilterBank(); - this->m_filterBankInitialised = true; - } - } - - bool MelSpectrogram::IsMelFilterBankInited() const - { - return this->m_filterBankInitialised; - } - - std::vector<float> MelSpectrogram::ComputeMelSpec(const std::vector<int16_t>& audioData, float trainingMean) - { - this->InitMelFilterBank(); - - /* TensorFlow way of normalizing .wav data to (-1, 1). */ - constexpr float normaliser = 1.0/(1<<15); - 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. */ - math::MathUtils::FftF32(this->m_frame, this->m_buffer, this->m_fftInstance); - - /* 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_err("Failed to apply MEL filter banks\n"); - } - - /* Convert to logarithmic scale */ - this->ConvertToLogarithmicScale(this->m_melEnergies); - - /* Perform mean subtraction. */ - for (auto& energy:this->m_melEnergies) { - energy -= trainingMean; - } - - return this->m_melEnergies; - } - - std::vector<std::vector<float>> MelSpectrogram::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 = MelSpectrogram::MelScale(this->m_params.m_melLoFreq, - this->m_params.m_useHtkMethod); - float melHighFreq = MelSpectrogram::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<uint32_t>(this->m_params.m_numFbankBins); - this->m_filterBankFilterLast = - std::vector<uint32_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; - - uint32_t firstIndex = 0; - uint32_t lastIndex = 0; - bool firstIndexFound = false; - 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 = MelSpectrogram::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 (!firstIndexFound) { - firstIndex = i; - firstIndexFound = true; - } - lastIndex = i; - } - } - - this->m_filterBankFilterFirst[bin] = firstIndex; - this->m_filterBankFilterLast[bin] = lastIndex; - - /* Copy the part we care about. */ - for (uint32_t i = firstIndex; i <= lastIndex; ++i) { - melFilterBank[bin].push_back(thisBin[i]); - } - } - - return melFilterBank; - } - -} /* namespace audio */ -} /* namespace app */ -} /* namespace arm */ |