diff options
Diffstat (limited to 'source/use_case/kws_asr/src')
| -rw-r--r-- | source/use_case/kws_asr/src/AsrClassifier.cc | 131 | ||||
| -rw-r--r-- | source/use_case/kws_asr/src/DsCnnModel.cc | 67 | ||||
| -rw-r--r-- | source/use_case/kws_asr/src/MainLoop.cc | 233 | ||||
| -rw-r--r-- | source/use_case/kws_asr/src/OutputDecode.cc | 47 | ||||
| -rw-r--r-- | source/use_case/kws_asr/src/UseCaseHandler.cc | 707 | ||||
| -rw-r--r-- | source/use_case/kws_asr/src/Wav2LetterMfcc.cc | 137 | ||||
| -rw-r--r-- | source/use_case/kws_asr/src/Wav2LetterModel.cc | 62 | ||||
| -rw-r--r-- | source/use_case/kws_asr/src/Wav2LetterPostprocess.cc | 155 | ||||
| -rw-r--r-- | source/use_case/kws_asr/src/Wav2LetterPreprocess.cc | 228 |
9 files changed, 1767 insertions, 0 deletions
diff --git a/source/use_case/kws_asr/src/AsrClassifier.cc b/source/use_case/kws_asr/src/AsrClassifier.cc new file mode 100644 index 0000000..bc86e09 --- /dev/null +++ b/source/use_case/kws_asr/src/AsrClassifier.cc @@ -0,0 +1,131 @@ +/* + * 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 "AsrClassifier.hpp" + +#include "hal.h" +#include "TensorFlowLiteMicro.hpp" +#include "Wav2LetterModel.hpp" + +template<typename T> +bool arm::app::AsrClassifier::_GetTopResults(TfLiteTensor* tensor, + std::vector<ClassificationResult>& vecResults, + const std::vector <std::string>& labels, double scale, double zeroPoint) +{ + const uint32_t nElems = tensor->dims->data[arm::app::Wav2LetterModel::ms_outputRowsIdx]; + const uint32_t nLetters = tensor->dims->data[arm::app::Wav2LetterModel::ms_outputColsIdx]; + + + /* NOTE: tensor's size verification against labels should be + * checked by the calling/public function. */ + if (nLetters < 1) { + return false; + } + + /* Final results' container. */ + vecResults = std::vector<ClassificationResult>(nElems); + + T* tensorData = tflite::GetTensorData<T>(tensor); + + /* Get the top 1 results. */ + for (uint32_t i = 0, row = 0; i < nElems; ++i, row+=nLetters) { + std::pair<T, uint32_t> top_1 = std::make_pair(tensorData[row + 0], 0); + + for (uint32_t j = 1; j < nLetters; ++j) { + if (top_1.first < tensorData[row + j]) { + top_1.first = tensorData[row + j]; + top_1.second = j; + } + } + + double score = static_cast<int> (top_1.first); + vecResults[i].m_normalisedVal = scale * (score - zeroPoint); + vecResults[i].m_label = labels[top_1.second]; + vecResults[i].m_labelIdx = top_1.second; + } + + return true; +} +template bool arm::app::AsrClassifier::_GetTopResults<uint8_t>(TfLiteTensor* tensor, + std::vector<ClassificationResult>& vecResults, + const std::vector <std::string>& labels, double scale, double zeroPoint); +template bool arm::app::AsrClassifier::_GetTopResults<int8_t>(TfLiteTensor* tensor, + std::vector<ClassificationResult>& vecResults, + const std::vector <std::string>& labels, double scale, double zeroPoint); + +bool arm::app::AsrClassifier::GetClassificationResults( + TfLiteTensor* outputTensor, + std::vector<ClassificationResult>& vecResults, + const std::vector <std::string>& labels, uint32_t topNCount) +{ + vecResults.clear(); + + constexpr int minTensorDims = static_cast<int>( + (arm::app::Wav2LetterModel::ms_outputRowsIdx > arm::app::Wav2LetterModel::ms_outputColsIdx)? + arm::app::Wav2LetterModel::ms_outputRowsIdx : arm::app::Wav2LetterModel::ms_outputColsIdx); + + constexpr uint32_t outColsIdx = arm::app::Wav2LetterModel::ms_outputColsIdx; + + /* Sanity checks. */ + if (outputTensor == nullptr) { + printf_err("Output vector is null pointer.\n"); + return false; + } else if (outputTensor->dims->size < minTensorDims) { + printf_err("Output tensor expected to be 3D (1, m, n)\n"); + return false; + } else if (static_cast<uint32_t>(outputTensor->dims->data[outColsIdx]) < topNCount) { + printf_err("Output vectors are smaller than %u\n", topNCount); + return false; + } else if (static_cast<uint32_t>(outputTensor->dims->data[outColsIdx]) != labels.size()) { + printf("Output size doesn't match the labels' size\n"); + return false; + } + + if (topNCount != 1) { + warn("TopNCount value ignored in this implementation\n"); + } + + /* To return the floating point values, we need quantization parameters. */ + QuantParams quantParams = GetTensorQuantParams(outputTensor); + + bool resultState; + + switch (outputTensor->type) { + case kTfLiteUInt8: + resultState = this->_GetTopResults<uint8_t>( + outputTensor, vecResults, + labels, quantParams.scale, + quantParams.offset); + break; + case kTfLiteInt8: + resultState = this->_GetTopResults<int8_t>( + outputTensor, vecResults, + labels, quantParams.scale, + quantParams.offset); + break; + default: + printf_err("Tensor type %s not supported by classifier\n", + TfLiteTypeGetName(outputTensor->type)); + return false; + } + + if (!resultState) { + printf_err("Failed to get sorted set\n"); + return false; + } + + return true; +}
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/DsCnnModel.cc b/source/use_case/kws_asr/src/DsCnnModel.cc new file mode 100644 index 0000000..b573a12 --- /dev/null +++ b/source/use_case/kws_asr/src/DsCnnModel.cc @@ -0,0 +1,67 @@ +/* + * 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 "DsCnnModel.hpp" + +#include "hal.h" + +namespace arm { +namespace app { +namespace kws { + extern uint8_t* GetModelPointer(); + extern size_t GetModelLen(); +} /* namespace kws */ +} /* namespace app */ +} /* namespace arm */ + +const tflite::MicroOpResolver& arm::app::DsCnnModel::GetOpResolver() +{ + return this->_m_opResolver; +} + +bool arm::app::DsCnnModel::EnlistOperations() +{ + this->_m_opResolver.AddAveragePool2D(); + this->_m_opResolver.AddConv2D(); + this->_m_opResolver.AddDepthwiseConv2D(); + this->_m_opResolver.AddFullyConnected(); + this->_m_opResolver.AddRelu(); + this->_m_opResolver.AddSoftmax(); + this->_m_opResolver.AddQuantize(); + this->_m_opResolver.AddDequantize(); + 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; +} + +const uint8_t* arm::app::DsCnnModel::ModelPointer() +{ + return arm::app::kws::GetModelPointer(); +} + +size_t arm::app::DsCnnModel::ModelSize() +{ + return arm::app::kws::GetModelLen(); +}
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/MainLoop.cc b/source/use_case/kws_asr/src/MainLoop.cc new file mode 100644 index 0000000..37146c9 --- /dev/null +++ b/source/use_case/kws_asr/src/MainLoop.cc @@ -0,0 +1,233 @@ +/* + * 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 "hal.h" /* Brings in platform definitions. */ +#include "InputFiles.hpp" /* For input images. */ +#include "Labels_dscnn.hpp" /* For DS-CNN label strings. */ +#include "Labels_wav2letter.hpp" /* For Wav2Letter label strings. */ +#include "Classifier.hpp" /* KWS classifier. */ +#include "AsrClassifier.hpp" /* ASR classifier. */ +#include "DsCnnModel.hpp" /* KWS model class for running inference. */ +#include "Wav2LetterModel.hpp" /* ASR model class for running inference. */ +#include "UseCaseCommonUtils.hpp" /* Utils functions. */ +#include "UseCaseHandler.hpp" /* Handlers for different user options. */ +#include "Wav2LetterPreprocess.hpp" /* ASR pre-processing class. */ +#include "Wav2LetterPostprocess.hpp"/* ASR post-processing class. */ + +using KwsClassifier = arm::app::Classifier; + +enum opcodes +{ + MENU_OPT_RUN_INF_NEXT = 1, /* Run on next vector. */ + MENU_OPT_RUN_INF_CHOSEN, /* Run on a user provided vector index. */ + MENU_OPT_RUN_INF_ALL, /* Run inference on all. */ + MENU_OPT_SHOW_MODEL_INFO, /* Show model info. */ + MENU_OPT_LIST_AUDIO_CLIPS /* List the current baked audio clips. */ +}; + +static void DisplayMenu() +{ + printf("\n\nUser input required\n"); + printf("Enter option number from:\n\n"); + printf(" %u. Classify next audio clip\n", MENU_OPT_RUN_INF_NEXT); + printf(" %u. Classify audio clip at chosen index\n", MENU_OPT_RUN_INF_CHOSEN); + printf(" %u. Run classification on all audio clips\n", MENU_OPT_RUN_INF_ALL); + printf(" %u. Show NN model info\n", MENU_OPT_SHOW_MODEL_INFO); + printf(" %u. List audio clips\n\n", MENU_OPT_LIST_AUDIO_CLIPS); + printf(" Choice: "); +} + +/** @brief Gets the number of MFCC features for a single window. */ +static uint32_t GetNumMfccFeatures(const arm::app::Model& model); + +/** @brief Gets the number of MFCC feature vectors to be computed. */ +static uint32_t GetNumMfccFeatureVectors(const arm::app::Model& model); + +/** @brief Gets the output context length (left and right) for post-processing. */ +static uint32_t GetOutputContextLen(const arm::app::Model& model, + uint32_t inputCtxLen); + +/** @brief Gets the output inner length for post-processing. */ +static uint32_t GetOutputInnerLen(const arm::app::Model& model, + uint32_t outputCtxLen); + +void main_loop(hal_platform& platform) +{ + /* Model wrapper objects. */ + arm::app::DsCnnModel kwsModel; + arm::app::Wav2LetterModel asrModel; + + /* Load the models. */ + if (!kwsModel.Init()) { + printf_err("Failed to initialise KWS model\n"); + return; + } + + /* Initialise the asr model using the same allocator from KWS + * to re-use the tensor arena. */ + if (!asrModel.Init(kwsModel.GetAllocator())) { + printf_err("Failed to initalise ASR model\n"); + return; + } + + /* Initialise ASR pre-processing. */ + arm::app::audio::asr::Preprocess prep( + GetNumMfccFeatures(asrModel), + arm::app::asr::g_FrameLength, + arm::app::asr::g_FrameStride, + GetNumMfccFeatureVectors(asrModel)); + + /* Initialise ASR post-processing. */ + const uint32_t outputCtxLen = GetOutputContextLen(asrModel, arm::app::asr::g_ctxLen); + const uint32_t blankTokenIdx = 28; + arm::app::audio::asr::Postprocess postp( + outputCtxLen, + GetOutputInnerLen(asrModel, outputCtxLen), + blankTokenIdx); + + /* Instantiate application context. */ + arm::app::ApplicationContext caseContext; + + caseContext.Set<hal_platform&>("platform", platform); + caseContext.Set<arm::app::Model&>("kwsmodel", kwsModel); + caseContext.Set<arm::app::Model&>("asrmodel", asrModel); + caseContext.Set<uint32_t>("clipIndex", 0); + caseContext.Set<uint32_t>("ctxLen", arm::app::asr::g_ctxLen); /* Left and right context length (MFCC feat vectors). */ + caseContext.Set<int>("kwsframeLength", arm::app::kws::g_FrameLength); + caseContext.Set<int>("kwsframeStride", arm::app::kws::g_FrameStride); + caseContext.Set<float>("kwsscoreThreshold", arm::app::kws::g_ScoreThreshold); /* Normalised score threshold. */ + caseContext.Set<uint32_t >("kwsNumMfcc", arm::app::kws::g_NumMfcc); + caseContext.Set<uint32_t >("kwsNumAudioWins", arm::app::kws::g_NumAudioWins); + + caseContext.Set<int>("asrframeLength", arm::app::asr::g_FrameLength); + caseContext.Set<int>("asrframeStride", arm::app::asr::g_FrameStride); + caseContext.Set<float>("asrscoreThreshold", arm::app::asr::g_ScoreThreshold); /* Normalised score threshold. */ + + KwsClassifier kwsClassifier; /* Classifier wrapper object. */ + arm::app::AsrClassifier asrClassifier; /* Classifier wrapper object. */ + caseContext.Set<arm::app::Classifier&>("kwsclassifier", kwsClassifier); + caseContext.Set<arm::app::AsrClassifier&>("asrclassifier", asrClassifier); + + caseContext.Set<arm::app::audio::asr::Preprocess&>("preprocess", prep); + caseContext.Set<arm::app::audio::asr::Postprocess&>("postprocess", postp); + + std::vector<std::string> asrLabels; + arm::app::asr::GetLabelsVector(asrLabels); + std::vector<std::string> kwsLabels; + arm::app::kws::GetLabelsVector(kwsLabels); + caseContext.Set<const std::vector <std::string>&>("asrlabels", asrLabels); + caseContext.Set<const std::vector <std::string>&>("kwslabels", kwsLabels); + + /* Index of the kws outputs we trigger ASR on. */ + caseContext.Set<uint32_t>("keywordindex", 2); + + /* Loop. */ + bool executionSuccessful = true; + constexpr bool bUseMenu = NUMBER_OF_FILES > 1 ? true : false; + + /* Loop. */ + do { + int menuOption = MENU_OPT_RUN_INF_NEXT; + if (bUseMenu) { + DisplayMenu(); + menuOption = arm::app::ReadUserInputAsInt(platform); + printf("\n"); + } + switch (menuOption) { + case MENU_OPT_RUN_INF_NEXT: + executionSuccessful = ClassifyAudioHandler( + caseContext, + caseContext.Get<uint32_t>("clipIndex"), + false); + break; + case MENU_OPT_RUN_INF_CHOSEN: { + printf(" Enter the audio clip index [0, %d]: ", + NUMBER_OF_FILES-1); + auto clipIndex = static_cast<uint32_t>( + arm::app::ReadUserInputAsInt(platform)); + executionSuccessful = ClassifyAudioHandler(caseContext, + clipIndex, + false); + break; + } + case MENU_OPT_RUN_INF_ALL: + executionSuccessful = ClassifyAudioHandler( + caseContext, + caseContext.Get<uint32_t>("clipIndex"), + true); + break; + case MENU_OPT_SHOW_MODEL_INFO: + executionSuccessful = kwsModel.ShowModelInfoHandler(); + executionSuccessful = asrModel.ShowModelInfoHandler(); + break; + case MENU_OPT_LIST_AUDIO_CLIPS: + executionSuccessful = ListFilesHandler(caseContext); + break; + default: + printf("Incorrect choice, try again."); + break; + } + } while (executionSuccessful && bUseMenu); + info("Main loop terminated.\n"); +} + +static uint32_t GetNumMfccFeatures(const arm::app::Model& model) +{ + TfLiteTensor* inputTensor = model.GetInputTensor(0); + const int inputCols = inputTensor->dims->data[arm::app::Wav2LetterModel::ms_inputColsIdx]; + if (0 != inputCols % 3) { + printf_err("Number of input columns is not a multiple of 3\n"); + } + return std::max(inputCols/3, 0); +} + +static uint32_t GetNumMfccFeatureVectors(const arm::app::Model& model) +{ + TfLiteTensor* inputTensor = model.GetInputTensor(0); + const int inputRows = inputTensor->dims->data[arm::app::Wav2LetterModel::ms_inputRowsIdx]; + return std::max(inputRows, 0); +} + +static uint32_t GetOutputContextLen(const arm::app::Model& model, const uint32_t inputCtxLen) +{ + const uint32_t inputRows = GetNumMfccFeatureVectors(model); + const uint32_t inputInnerLen = inputRows - (2 * inputCtxLen); + constexpr uint32_t ms_outputRowsIdx = arm::app::Wav2LetterModel::ms_outputRowsIdx; + + /* Check to make sure that the input tensor supports the above context and inner lengths. */ + if (inputRows <= 2 * inputCtxLen || inputRows <= inputInnerLen) { + printf_err("Input rows not compatible with ctx of %u\n", + inputCtxLen); + return 0; + } + + TfLiteTensor* outputTensor = model.GetOutputTensor(0); + const uint32_t outputRows = std::max(outputTensor->dims->data[ms_outputRowsIdx], 0); + + const float tensorColRatio = static_cast<float>(inputRows)/ + static_cast<float>(outputRows); + + return std::round(static_cast<float>(inputCtxLen)/tensorColRatio); +} + +static uint32_t GetOutputInnerLen(const arm::app::Model& model, + const uint32_t outputCtxLen) +{ + constexpr uint32_t ms_outputRowsIdx = arm::app::Wav2LetterModel::ms_outputRowsIdx; + TfLiteTensor* outputTensor = model.GetOutputTensor(0); + const uint32_t outputRows = std::max(outputTensor->dims->data[ms_outputRowsIdx], 0); + return (outputRows - (2 * outputCtxLen)); +} diff --git a/source/use_case/kws_asr/src/OutputDecode.cc b/source/use_case/kws_asr/src/OutputDecode.cc new file mode 100644 index 0000000..41fbe07 --- /dev/null +++ b/source/use_case/kws_asr/src/OutputDecode.cc @@ -0,0 +1,47 @@ +/* + * 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 "OutputDecode.hpp" + +namespace arm { +namespace app { +namespace audio { +namespace asr { + + std::string DecodeOutput(const std::vector<ClassificationResult>& vecResults) + { + std::string CleanOutputBuffer; + + for (size_t i = 0; i < vecResults.size(); ++i) /* For all elements in vector. */ + { + while (i+1 < vecResults.size() && + vecResults[i].m_label == vecResults[i+1].m_label) /* While the current element is equal to the next, ignore it and move on. */ + { + ++i; + } + if (vecResults[i].m_label != "$") /* $ is a character used to represent unknown and double characters so should not be in output. */ + { + CleanOutputBuffer += vecResults[i].m_label; /* If the element is different to the next, it will be appended to CleanOutputBuffer. */ + } + } + + return CleanOutputBuffer; /* Return string type containing clean output. */ + } + +} /* namespace asr */ +} /* namespace audio */ +} /* namespace app */ +} /* namespace arm */ diff --git a/source/use_case/kws_asr/src/UseCaseHandler.cc b/source/use_case/kws_asr/src/UseCaseHandler.cc new file mode 100644 index 0000000..c50796f --- /dev/null +++ b/source/use_case/kws_asr/src/UseCaseHandler.cc @@ -0,0 +1,707 @@ +/* + * 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 "UseCaseHandler.hpp" + +#include "hal.h" +#include "InputFiles.hpp" +#include "AudioUtils.hpp" +#include "UseCaseCommonUtils.hpp" +#include "DsCnnModel.hpp" +#include "DsCnnMfcc.hpp" +#include "Classifier.hpp" +#include "KwsResult.hpp" +#include "Wav2LetterMfcc.hpp" +#include "Wav2LetterPreprocess.hpp" +#include "Wav2LetterPostprocess.hpp" +#include "AsrResult.hpp" +#include "AsrClassifier.hpp" +#include "OutputDecode.hpp" + + +using KwsClassifier = arm::app::Classifier; + +namespace arm { +namespace app { + + enum AsrOutputReductionAxis { + AxisRow = 1, + AxisCol = 2 + }; + + struct KWSOutput { + bool executionSuccess = false; + const int16_t* asrAudioStart = nullptr; + int32_t asrAudioSamples = 0; + }; + + /** + * @brief Helper function to increment current audio clip index + * @param[in,out] ctx pointer to the application context object + **/ + static void _IncrementAppCtxClipIdx(ApplicationContext& ctx); + + /** + * @brief Helper function to increment current audio clip index + * @param[in,out] ctx pointer to the application context object + **/ + static void _IncrementAppCtxClipIdx(ApplicationContext& ctx); + + /** + * @brief Helper function to set the audio clip index + * @param[in,out] ctx pointer to the application context object + * @param[in] idx value to be set + * @return true if index is set, false otherwise + **/ + static bool _SetAppCtxClipIdx(ApplicationContext& ctx, uint32_t idx); + + /** + * @brief Presents kws inference results using the data presentation + * object. + * @param[in] platform reference to the hal platform object + * @param[in] results vector of classification results to be displayed + * @param[in] infTimeMs inference time in milliseconds, if available + * Otherwise, this can be passed in as 0. + * @return true if successful, false otherwise + **/ + static bool _PresentInferenceResult(hal_platform& platform, std::vector<arm::app::kws::KwsResult>& results); + + /** + * @brief Presents asr inference results using the data presentation + * object. + * @param[in] platform reference to the hal platform object + * @param[in] results vector of classification results to be displayed + * @param[in] infTimeMs inference time in milliseconds, if available + * Otherwise, this can be passed in as 0. + * @return true if successful, false otherwise + **/ + static bool _PresentInferenceResult(hal_platform& platform, std::vector<arm::app::asr::AsrResult>& results); + + /** + * @brief Returns a function to perform feature calculation and populates input tensor data with + * MFCC data. + * + * Input tensor data type check is performed to choose correct MFCC feature data type. + * If tensor has an integer data type then original features are quantised. + * + * Warning: mfcc calculator provided as input must have the same life scope as returned function. + * + * @param[in] mfcc MFCC feature calculator. + * @param[in,out] inputTensor Input tensor pointer to store calculated features. + * @param[in] cacheSize Size of the feture vectors cache (number of feature vectors). + * + * @return function function to be called providing audio sample and sliding window index. + **/ + static std::function<void (std::vector<int16_t>&, int, bool, size_t)> + GetFeatureCalculator(audio::DsCnnMFCC& mfcc, + TfLiteTensor* inputTensor, + size_t cacheSize); + + /** + * @brief Performs the KWS pipeline. + * @param[in,out] ctx pointer to the application context object + * + * @return KWSOutput struct containing pointer to audio data where ASR should begin + * and how much data to process. + */ + static KWSOutput doKws(ApplicationContext& ctx) { + constexpr uint32_t dataPsnTxtInfStartX = 20; + constexpr uint32_t dataPsnTxtInfStartY = 40; + + constexpr int minTensorDims = static_cast<int>( + (arm::app::DsCnnModel::ms_inputRowsIdx > arm::app::DsCnnModel::ms_inputColsIdx)? + arm::app::DsCnnModel::ms_inputRowsIdx : arm::app::DsCnnModel::ms_inputColsIdx); + + KWSOutput output; + + auto& kwsModel = ctx.Get<Model&>("kwsmodel"); + if (!kwsModel.IsInited()) { + printf_err("KWS model has not been initialised\n"); + return output; + } + + const int kwsFrameLength = ctx.Get<int>("kwsframeLength"); + const int kwsFrameStride = ctx.Get<int>("kwsframeStride"); + const float kwsScoreThreshold = ctx.Get<float>("kwsscoreThreshold"); + + TfLiteTensor* kwsOutputTensor = kwsModel.GetOutputTensor(0); + TfLiteTensor* kwsInputTensor = kwsModel.GetInputTensor(0); + + if (!kwsInputTensor->dims) { + printf_err("Invalid input tensor dims\n"); + return output; + } else if (kwsInputTensor->dims->size < minTensorDims) { + printf_err("Input tensor dimension should be >= %d\n", minTensorDims); + return output; + } + + const uint32_t kwsNumMfccFeats = ctx.Get<uint32_t>("kwsNumMfcc"); + const uint32_t kwsNumAudioWindows = ctx.Get<uint32_t>("kwsNumAudioWins"); + + audio::DsCnnMFCC kwsMfcc = audio::DsCnnMFCC(kwsNumMfccFeats, kwsFrameLength); + kwsMfcc.Init(); + + /* Deduce the data length required for 1 KWS inference from the network parameters. */ + auto kwsAudioDataWindowSize = kwsNumAudioWindows * kwsFrameStride + + (kwsFrameLength - kwsFrameStride); + auto kwsMfccWindowSize = kwsFrameLength; + auto kwsMfccWindowStride = kwsFrameStride; + + /* We are choosing to move by half the window size => for a 1 second window size, + * this means an overlap of 0.5 seconds. */ + auto kwsAudioDataStride = kwsAudioDataWindowSize / 2; + + info("KWS audio data window size %u\n", kwsAudioDataWindowSize); + + /* Stride must be multiple of mfcc features window stride to re-use features. */ + if (0 != kwsAudioDataStride % kwsMfccWindowStride) { + kwsAudioDataStride -= kwsAudioDataStride % kwsMfccWindowStride; + } + + auto kwsMfccVectorsInAudioStride = kwsAudioDataStride/kwsMfccWindowStride; + + /* We expect to be sampling 1 second worth of data at a time + * NOTE: This is only used for time stamp calculation. */ + const float kwsAudioParamsSecondsPerSample = 1.0/audio::DsCnnMFCC::ms_defaultSamplingFreq; + + auto currentIndex = ctx.Get<uint32_t>("clipIndex"); + + /* Creating a mfcc features sliding window for the data required for 1 inference. */ + auto kwsAudioMFCCWindowSlider = audio::SlidingWindow<const int16_t>( + get_audio_array(currentIndex), + kwsAudioDataWindowSize, kwsMfccWindowSize, + kwsMfccWindowStride); + + /* Creating a sliding window through the whole audio clip. */ + auto audioDataSlider = audio::SlidingWindow<const int16_t>( + get_audio_array(currentIndex), + get_audio_array_size(currentIndex), + kwsAudioDataWindowSize, kwsAudioDataStride); + + /* Calculate number of the feature vectors in the window overlap region. + * These feature vectors will be reused.*/ + size_t numberOfReusedFeatureVectors = kwsAudioMFCCWindowSlider.TotalStrides() + 1 + - kwsMfccVectorsInAudioStride; + + auto kwsMfccFeatureCalc = GetFeatureCalculator(kwsMfcc, kwsInputTensor, + numberOfReusedFeatureVectors); + + if (!kwsMfccFeatureCalc){ + return output; + } + + /* Container for KWS results. */ + std::vector<arm::app::kws::KwsResult> kwsResults; + + /* Display message on the LCD - inference running. */ + auto& platform = ctx.Get<hal_platform&>("platform"); + std::string str_inf{"Running KWS inference... "}; + platform.data_psn->present_data_text( + str_inf.c_str(), str_inf.size(), + dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); + + info("Running KWS inference on audio clip %u => %s\n", + currentIndex, get_filename(currentIndex)); + + /* Start sliding through audio clip. */ + while (audioDataSlider.HasNext()) { + const int16_t* inferenceWindow = audioDataSlider.Next(); + + /* We moved to the next window - set the features sliding to the new address. */ + kwsAudioMFCCWindowSlider.Reset(inferenceWindow); + + /* The first window does not have cache ready. */ + bool useCache = audioDataSlider.Index() > 0 && numberOfReusedFeatureVectors > 0; + + /* Start calculating features inside one audio sliding window. */ + while (kwsAudioMFCCWindowSlider.HasNext()) { + const int16_t* kwsMfccWindow = kwsAudioMFCCWindowSlider.Next(); + std::vector<int16_t> kwsMfccAudioData = + std::vector<int16_t>(kwsMfccWindow, kwsMfccWindow + kwsMfccWindowSize); + + /* Compute features for this window and write them to input tensor. */ + kwsMfccFeatureCalc(kwsMfccAudioData, + kwsAudioMFCCWindowSlider.Index(), + useCache, + kwsMfccVectorsInAudioStride); + } + + info("Inference %zu/%zu\n", audioDataSlider.Index() + 1, + audioDataSlider.TotalStrides() + 1); + + /* Run inference over this audio clip sliding window. */ + arm::app::RunInference(platform, kwsModel); + + std::vector<ClassificationResult> kwsClassificationResult; + auto& kwsClassifier = ctx.Get<KwsClassifier&>("kwsclassifier"); + + kwsClassifier.GetClassificationResults( + kwsOutputTensor, kwsClassificationResult, + ctx.Get<std::vector<std::string>&>("kwslabels"), 1); + + kwsResults.emplace_back( + kws::KwsResult( + kwsClassificationResult, + audioDataSlider.Index() * kwsAudioParamsSecondsPerSample * kwsAudioDataStride, + audioDataSlider.Index(), kwsScoreThreshold) + ); + + /* Keyword detected. */ + if (kwsClassificationResult[0].m_labelIdx == ctx.Get<uint32_t>("keywordindex")) { + output.asrAudioStart = inferenceWindow + kwsAudioDataWindowSize; + output.asrAudioSamples = get_audio_array_size(currentIndex) - + (audioDataSlider.NextWindowStartIndex() - + kwsAudioDataStride + kwsAudioDataWindowSize); + break; + } + +#if VERIFY_TEST_OUTPUT + arm::app::DumpTensor(kwsOutputTensor); +#endif /* VERIFY_TEST_OUTPUT */ + + } /* while (audioDataSlider.HasNext()) */ + + /* Erase. */ + str_inf = std::string(str_inf.size(), ' '); + platform.data_psn->present_data_text( + str_inf.c_str(), str_inf.size(), + dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); + + if (!_PresentInferenceResult(platform, kwsResults)) { + return output; + } + + output.executionSuccess = true; + return output; + } + + /** + * @brief Performs the ASR pipeline. + * + * @param ctx[in/out] pointer to the application context object + * @param kwsOutput[in] struct containing pointer to audio data where ASR should begin + * and how much data to process + * @return bool true if pipeline executed without failure + */ + static bool doAsr(ApplicationContext& ctx, const KWSOutput& kwsOutput) { + constexpr uint32_t dataPsnTxtInfStartX = 20; + constexpr uint32_t dataPsnTxtInfStartY = 40; + + auto& platform = ctx.Get<hal_platform&>("platform"); + platform.data_psn->clear(COLOR_BLACK); + + /* Get model reference. */ + auto& asrModel = ctx.Get<Model&>("asrmodel"); + if (!asrModel.IsInited()) { + printf_err("ASR model has not been initialised\n"); + return false; + } + + /* Get score threshold to be applied for the classifier (post-inference). */ + auto asrScoreThreshold = ctx.Get<float>("asrscoreThreshold"); + + /* Dimensions of the tensor should have been verified by the callee. */ + TfLiteTensor* asrInputTensor = asrModel.GetInputTensor(0); + TfLiteTensor* asrOutputTensor = asrModel.GetOutputTensor(0); + const uint32_t asrInputRows = asrInputTensor->dims->data[arm::app::Wav2LetterModel::ms_inputRowsIdx]; + + /* Populate ASR MFCC related parameters. */ + auto asrMfccParamsWinLen = ctx.Get<uint32_t>("asrframeLength"); + auto asrMfccParamsWinStride = ctx.Get<uint32_t>("asrframeStride"); + + /* Populate ASR inference context and inner lengths for input. */ + auto asrInputCtxLen = ctx.Get<uint32_t>("ctxLen"); + const uint32_t asrInputInnerLen = asrInputRows - (2 * asrInputCtxLen); + + /* Make sure the input tensor supports the above context and inner lengths. */ + if (asrInputRows <= 2 * asrInputCtxLen || asrInputRows <= asrInputInnerLen) { + printf_err("ASR input rows not compatible with ctx length %u\n", asrInputCtxLen); + return false; + } + + /* Audio data stride corresponds to inputInnerLen feature vectors. */ + const uint32_t asrAudioParamsWinLen = (asrInputRows - 1) * + asrMfccParamsWinStride + (asrMfccParamsWinLen); + const uint32_t asrAudioParamsWinStride = asrInputInnerLen * asrMfccParamsWinStride; + const float asrAudioParamsSecondsPerSample = + (1.0/audio::Wav2LetterMFCC::ms_defaultSamplingFreq); + + /* Get pre/post-processing objects */ + auto& asrPrep = ctx.Get<audio::asr::Preprocess&>("preprocess"); + auto& asrPostp = ctx.Get<audio::asr::Postprocess&>("postprocess"); + + /* Set default reduction axis for post-processing. */ + const uint32_t reductionAxis = arm::app::Wav2LetterModel::ms_outputRowsIdx; + + /* Get the remaining audio buffer and respective size from KWS results. */ + const int16_t* audioArr = kwsOutput.asrAudioStart; + const uint32_t audioArrSize = kwsOutput.asrAudioSamples; + + /* Audio clip must have enough samples to produce 1 MFCC feature. */ + std::vector<int16_t> audioBuffer = std::vector<int16_t>(audioArr, audioArr + audioArrSize); + if (audioArrSize < asrMfccParamsWinLen) { + printf_err("Not enough audio samples, minimum needed is %u\n", asrMfccParamsWinLen); + return false; + } + + /* Initialise an audio slider. */ + auto audioDataSlider = audio::ASRSlidingWindow<const int16_t>( + audioBuffer.data(), + audioBuffer.size(), + asrAudioParamsWinLen, + asrAudioParamsWinStride); + + /* Declare a container for results. */ + std::vector<arm::app::asr::AsrResult> asrResults; + + /* Display message on the LCD - inference running. */ + std::string str_inf{"Running ASR inference... "}; + platform.data_psn->present_data_text( + str_inf.c_str(), str_inf.size(), + dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); + + size_t asrInferenceWindowLen = asrAudioParamsWinLen; + + /* Start sliding through audio clip. */ + while (audioDataSlider.HasNext()) { + + /* If not enough audio see how much can be sent for processing. */ + size_t nextStartIndex = audioDataSlider.NextWindowStartIndex(); + if (nextStartIndex + asrAudioParamsWinLen > audioBuffer.size()) { + asrInferenceWindowLen = audioBuffer.size() - nextStartIndex; + } + + const int16_t* asrInferenceWindow = audioDataSlider.Next(); + + info("Inference %zu/%zu\n", audioDataSlider.Index() + 1, + static_cast<size_t>(ceilf(audioDataSlider.FractionalTotalStrides() + 1))); + + Profiler prepProfiler{&platform, "pre-processing"}; + prepProfiler.StartProfiling(); + + /* Calculate MFCCs, deltas and populate the input tensor. */ + asrPrep.Invoke(asrInferenceWindow, asrInferenceWindowLen, asrInputTensor); + + prepProfiler.StopProfiling(); + std::string prepProfileResults = prepProfiler.GetResultsAndReset(); + info("%s\n", prepProfileResults.c_str()); + + /* Run inference over this audio clip sliding window. */ + arm::app::RunInference(platform, asrModel); + + /* Post-process. */ + asrPostp.Invoke(asrOutputTensor, reductionAxis, !audioDataSlider.HasNext()); + + /* Get results. */ + std::vector<ClassificationResult> asrClassificationResult; + auto& asrClassifier = ctx.Get<AsrClassifier&>("asrclassifier"); + asrClassifier.GetClassificationResults( + asrOutputTensor, asrClassificationResult, + ctx.Get<std::vector<std::string>&>("asrlabels"), 1); + + asrResults.emplace_back(asr::AsrResult(asrClassificationResult, + (audioDataSlider.Index() * + asrAudioParamsSecondsPerSample * + asrAudioParamsWinStride), + audioDataSlider.Index(), asrScoreThreshold)); + +#if VERIFY_TEST_OUTPUT + arm::app::DumpTensor(asrOutputTensor, asrOutputTensor->dims->data[arm::app::Wav2LetterModel::ms_outputColsIdx]); +#endif /* VERIFY_TEST_OUTPUT */ + + /* Erase */ + str_inf = std::string(str_inf.size(), ' '); + platform.data_psn->present_data_text( + str_inf.c_str(), str_inf.size(), + dataPsnTxtInfStartX, dataPsnTxtInfStartY, false); + } + if (!_PresentInferenceResult(platform, asrResults)) { + return false; + } + + return true; + } + + /* Audio inference classification handler. */ + bool ClassifyAudioHandler(ApplicationContext& ctx, uint32_t clipIndex, bool runAll) + { + auto& platform = ctx.Get<hal_platform&>("platform"); + platform.data_psn->clear(COLOR_BLACK); + + /* If the request has a valid size, set the audio index. */ + if (clipIndex < NUMBER_OF_FILES) { + if (!_SetAppCtxClipIdx(ctx, clipIndex)) { + return false; + } + } + + auto startClipIdx = ctx.Get<uint32_t>("clipIndex"); + + do { + KWSOutput kwsOutput = doKws(ctx); + if (!kwsOutput.executionSuccess) { + return false; + } + + if (kwsOutput.asrAudioStart != nullptr && kwsOutput.asrAudioSamples > 0) { + info("Keyword spotted\n"); + if(!doAsr(ctx, kwsOutput)) { + printf_err("ASR failed"); + return false; + } + } + + _IncrementAppCtxClipIdx(ctx); + + } while (runAll && ctx.Get<uint32_t>("clipIndex") != startClipIdx); + + return true; + } + + static void _IncrementAppCtxClipIdx(ApplicationContext& ctx) + { + auto curAudioIdx = ctx.Get<uint32_t>("clipIndex"); + + if (curAudioIdx + 1 >= NUMBER_OF_FILES) { + ctx.Set<uint32_t>("clipIndex", 0); + return; + } + ++curAudioIdx; + ctx.Set<uint32_t>("clipIndex", curAudioIdx); + } + + static bool _SetAppCtxClipIdx(ApplicationContext& ctx, const uint32_t idx) + { + if (idx >= NUMBER_OF_FILES) { + printf_err("Invalid idx %u (expected less than %u)\n", + idx, NUMBER_OF_FILES); + return false; + } + ctx.Set<uint32_t>("clipIndex", idx); + return true; + } + + static bool _PresentInferenceResult(hal_platform& platform, + std::vector<arm::app::kws::KwsResult>& results) + { + constexpr uint32_t dataPsnTxtStartX1 = 20; + constexpr uint32_t dataPsnTxtStartY1 = 30; + constexpr uint32_t dataPsnTxtYIncr = 16; /* Row index increment. */ + + platform.data_psn->set_text_color(COLOR_GREEN); + + /* Display each result. */ + uint32_t rowIdx1 = dataPsnTxtStartY1 + 2 * dataPsnTxtYIncr; + + for (uint32_t i = 0; i < results.size(); ++i) { + + std::string topKeyword{"<none>"}; + float score = 0.f; + + if (results[i].m_resultVec.size()) { + topKeyword = results[i].m_resultVec[0].m_label; + score = results[i].m_resultVec[0].m_normalisedVal; + } + + std::string resultStr = + std::string{"@"} + std::to_string(results[i].m_timeStamp) + + std::string{"s: "} + topKeyword + std::string{" ("} + + std::to_string(static_cast<int>(score * 100)) + std::string{"%)"}; + + platform.data_psn->present_data_text( + resultStr.c_str(), resultStr.size(), + dataPsnTxtStartX1, rowIdx1, 0); + rowIdx1 += dataPsnTxtYIncr; + + info("For timestamp: %f (inference #: %u); threshold: %f\n", + results[i].m_timeStamp, results[i].m_inferenceNumber, + results[i].m_threshold); + for (uint32_t j = 0; j < results[i].m_resultVec.size(); ++j) { + info("\t\tlabel @ %u: %s, score: %f\n", j, + results[i].m_resultVec[j].m_label.c_str(), + results[i].m_resultVec[j].m_normalisedVal); + } + } + + return true; + } + + static bool _PresentInferenceResult(hal_platform& platform, std::vector<arm::app::asr::AsrResult>& results) + { + constexpr uint32_t dataPsnTxtStartX1 = 20; + constexpr uint32_t dataPsnTxtStartY1 = 80; + constexpr bool allow_multiple_lines = true; + + platform.data_psn->set_text_color(COLOR_GREEN); + + /* Results from multiple inferences should be combined before processing. */ + std::vector<arm::app::ClassificationResult> combinedResults; + for (auto& result : results) { + combinedResults.insert(combinedResults.end(), + result.m_resultVec.begin(), + result.m_resultVec.end()); + } + + for (auto& result : results) { + /* Get the final result string using the decoder. */ + std::string infResultStr = audio::asr::DecodeOutput(result.m_resultVec); + + info("Result for inf %u: %s\n", result.m_inferenceNumber, + infResultStr.c_str()); + } + + std::string finalResultStr = audio::asr::DecodeOutput(combinedResults); + + platform.data_psn->present_data_text( + finalResultStr.c_str(), finalResultStr.size(), + dataPsnTxtStartX1, dataPsnTxtStartY1, allow_multiple_lines); + + info("Final result: %s\n", finalResultStr.c_str()); + return true; + } + + /** + * @brief Generic feature calculator factory. + * + * Returns lambda function to compute features using features cache. + * Real features math is done by a lambda function provided as a parameter. + * Features are written to input tensor memory. + * + * @tparam T feature vector type. + * @param inputTensor model input tensor pointer. + * @param cacheSize number of feature vectors to cache. Defined by the sliding window overlap. + * @param compute features calculator function. + * @return lambda function to compute features. + **/ + template<class T> + std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> + _FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, + std::function<std::vector<T> (std::vector<int16_t>& )> compute) + { + /* Feature cache to be captured by lambda function. */ + static std::vector<std::vector<T>> featureCache = std::vector<std::vector<T>>(cacheSize); + + return [=](std::vector<int16_t>& audioDataWindow, + size_t index, + bool useCache, + size_t featuresOverlapIndex) + { + T* tensorData = tflite::GetTensorData<T>(inputTensor); + std::vector<T> features; + + /* Reuse features from cache if cache is ready and sliding windows overlap. + * Overlap is in the beginning of sliding window with a size of a feature cache. + */ + if (useCache && index < featureCache.size()) { + features = std::move(featureCache[index]); + } else { + features = std::move(compute(audioDataWindow)); + } + auto size = features.size(); + auto sizeBytes = sizeof(T) * size; + std::memcpy(tensorData + (index * size), features.data(), sizeBytes); + + /* Start renewing cache as soon iteration goes out of the windows overlap. */ + if (index >= featuresOverlapIndex) { + featureCache[index - featuresOverlapIndex] = std::move(features); + } + }; + } + + template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> + _FeatureCalc<int8_t>(TfLiteTensor* inputTensor, + size_t cacheSize, + std::function<std::vector<int8_t> (std::vector<int16_t>& )> compute); + + template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> + _FeatureCalc<uint8_t>(TfLiteTensor* inputTensor, + size_t cacheSize, + std::function<std::vector<uint8_t> (std::vector<int16_t>& )> compute); + + template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> + _FeatureCalc<int16_t>(TfLiteTensor* inputTensor, + size_t cacheSize, + std::function<std::vector<int16_t> (std::vector<int16_t>& )> compute); + + template std::function<void(std::vector<int16_t>&, size_t, bool, size_t)> + _FeatureCalc<float>(TfLiteTensor* inputTensor, + size_t cacheSize, + std::function<std::vector<float>(std::vector<int16_t>&)> compute); + + + static std::function<void (std::vector<int16_t>&, int, bool, size_t)> + GetFeatureCalculator(audio::DsCnnMFCC& mfcc, TfLiteTensor* inputTensor, size_t cacheSize) + { + std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> mfccFeatureCalc; + + TfLiteQuantization quant = inputTensor->quantization; + + if (kTfLiteAffineQuantization == quant.type) { + + auto* quantParams = (TfLiteAffineQuantization*) quant.params; + const float quantScale = quantParams->scale->data[0]; + const int quantOffset = quantParams->zero_point->data[0]; + + switch (inputTensor->type) { + case kTfLiteInt8: { + mfccFeatureCalc = _FeatureCalc<int8_t>(inputTensor, + cacheSize, + [=, &mfcc](std::vector<int16_t>& audioDataWindow) { + return mfcc.MfccComputeQuant<int8_t>(audioDataWindow, + quantScale, + quantOffset); + } + ); + break; + } + case kTfLiteUInt8: { + mfccFeatureCalc = _FeatureCalc<uint8_t>(inputTensor, + cacheSize, + [=, &mfcc](std::vector<int16_t>& audioDataWindow) { + return mfcc.MfccComputeQuant<uint8_t>(audioDataWindow, + quantScale, + quantOffset); + } + ); + break; + } + case kTfLiteInt16: { + mfccFeatureCalc = _FeatureCalc<int16_t>(inputTensor, + cacheSize, + [=, &mfcc](std::vector<int16_t>& audioDataWindow) { + return mfcc.MfccComputeQuant<int16_t>(audioDataWindow, + quantScale, + quantOffset); + } + ); + break; + } + default: + printf_err("Tensor type %s not supported\n", TfLiteTypeGetName(inputTensor->type)); + } + + + } else { + mfccFeatureCalc = mfccFeatureCalc = _FeatureCalc<float>(inputTensor, + cacheSize, + [&mfcc](std::vector<int16_t>& audioDataWindow) { + return mfcc.MfccCompute(audioDataWindow); + }); + } + return mfccFeatureCalc; + } +} /* namespace app */ +} /* namespace arm */
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/Wav2LetterMfcc.cc b/source/use_case/kws_asr/src/Wav2LetterMfcc.cc new file mode 100644 index 0000000..80e4a26 --- /dev/null +++ b/source/use_case/kws_asr/src/Wav2LetterMfcc.cc @@ -0,0 +1,137 @@ +/* + * 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 "Wav2LetterMfcc.hpp" + +#include "PlatformMath.hpp" + +#include <cfloat> + +namespace arm { +namespace app { +namespace audio { + + bool Wav2LetterMFCC::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_err("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, same value used in librosa. */ + 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 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. */ + math::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); + } + } + + 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 * math::MathUtils::SqrtF32(1.0f / + static_cast<float>(4*inputLength)); + const float normalizer = 2 * math::MathUtils::SqrtF32(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 * + math::MathUtils::CosineF32((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))); + } + +} /* namespace audio */ +} /* namespace app */ +} /* namespace arm */ diff --git a/source/use_case/kws_asr/src/Wav2LetterModel.cc b/source/use_case/kws_asr/src/Wav2LetterModel.cc new file mode 100644 index 0000000..2114a3f --- /dev/null +++ b/source/use_case/kws_asr/src/Wav2LetterModel.cc @@ -0,0 +1,62 @@ +/* + * 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 "Wav2LetterModel.hpp" + +#include "hal.h" + +namespace arm { +namespace app { +namespace asr { + extern uint8_t* GetModelPointer(); + extern size_t GetModelLen(); +} +} /* namespace app */ +} /* namespace arm */ + +const tflite::MicroOpResolver& arm::app::Wav2LetterModel::GetOpResolver() +{ + return this->_m_opResolver; +} + +bool arm::app::Wav2LetterModel::EnlistOperations() +{ + this->_m_opResolver.AddConv2D(); + this->_m_opResolver.AddMul(); + this->_m_opResolver.AddMaximum(); + 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; +} + +const uint8_t* arm::app::Wav2LetterModel::ModelPointer() +{ + return arm::app::asr::GetModelPointer(); +} + +size_t arm::app::Wav2LetterModel::ModelSize() +{ + return arm::app::asr::GetModelLen(); +}
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/Wav2LetterPostprocess.cc b/source/use_case/kws_asr/src/Wav2LetterPostprocess.cc new file mode 100644 index 0000000..b173968 --- /dev/null +++ b/source/use_case/kws_asr/src/Wav2LetterPostprocess.cc @@ -0,0 +1,155 @@ +/* + * 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 "Wav2LetterPostprocess.hpp" + +#include "Wav2LetterModel.hpp" + +namespace arm { +namespace app { +namespace audio { +namespace asr { + + Postprocess::Postprocess(const uint32_t contextLen, + const uint32_t innerLen, + const uint32_t blankTokenIdx) + : _m_contextLen(contextLen), + _m_innerLen(innerLen), + _m_totalLen(2 * this->_m_contextLen + this->_m_innerLen), + _m_countIterations(0), + _m_blankTokenIdx(blankTokenIdx) + {} + + bool Postprocess::Invoke(TfLiteTensor* tensor, + const uint32_t axisIdx, + const bool lastIteration) + { + /* Basic checks. */ + if (!this->_IsInputValid(tensor, axisIdx)) { + return false; + } + + /* Irrespective of tensor type, we use unsigned "byte" */ + uint8_t* ptrData = tflite::GetTensorData<uint8_t>(tensor); + const uint32_t elemSz = this->_GetTensorElementSize(tensor); + + /* Other sanity checks. */ + if (0 == elemSz) { + printf_err("Tensor type not supported for post processing\n"); + return false; + } else if (elemSz * this->_m_totalLen > tensor->bytes) { + printf_err("Insufficient number of tensor bytes\n"); + return false; + } + + /* Which axis do we need to process? */ + switch (axisIdx) { + case arm::app::Wav2LetterModel::ms_outputRowsIdx: + return this->_EraseSectionsRowWise(ptrData, + elemSz * tensor->dims->data[arm::app::Wav2LetterModel::ms_outputColsIdx], + lastIteration); + default: + printf_err("Unsupported axis index: %u\n", axisIdx); + } + + return false; + } + + bool Postprocess::_IsInputValid(TfLiteTensor* tensor, + const uint32_t axisIdx) const + { + if (nullptr == tensor) { + return false; + } + + if (static_cast<int>(axisIdx) >= tensor->dims->size) { + printf_err("Invalid axis index: %u; Max: %d\n", + axisIdx, tensor->dims->size); + return false; + } + + if (static_cast<int>(this->_m_totalLen) != + tensor->dims->data[axisIdx]) { + printf_err("Unexpected tensor dimension for axis %d, \n", + tensor->dims->data[axisIdx]); + return false; + } + + return true; + } + + uint32_t Postprocess::_GetTensorElementSize(TfLiteTensor* tensor) + { + switch(tensor->type) { + case kTfLiteUInt8: + return 1; + case kTfLiteInt8: + return 1; + case kTfLiteInt16: + return 2; + case kTfLiteInt32: + return 4; + case kTfLiteFloat32: + return 4; + default: + printf_err("Unsupported tensor type %s\n", + TfLiteTypeGetName(tensor->type)); + } + + return 0; + } + + bool Postprocess::_EraseSectionsRowWise( + uint8_t* ptrData, + const uint32_t strideSzBytes, + const bool lastIteration) + { + /* In this case, the "zero-ing" is quite simple as the region + * to be zeroed sits in contiguous memory (row-major). */ + const uint32_t eraseLen = strideSzBytes * this->_m_contextLen; + + /* Erase left context? */ + if (this->_m_countIterations > 0) { + /* Set output of each classification window to the blank token. */ + std::memset(ptrData, 0, eraseLen); + for (size_t windowIdx = 0; windowIdx < this->_m_contextLen; windowIdx++) { + ptrData[windowIdx*strideSzBytes + this->_m_blankTokenIdx] = 1; + } + } + + /* Erase right context? */ + if (false == lastIteration) { + uint8_t * rightCtxPtr = ptrData + (strideSzBytes * (this->_m_contextLen + this->_m_innerLen)); + /* Set output of each classification window to the blank token. */ + std::memset(rightCtxPtr, 0, eraseLen); + for (size_t windowIdx = 0; windowIdx < this->_m_contextLen; windowIdx++) { + rightCtxPtr[windowIdx*strideSzBytes + this->_m_blankTokenIdx] = 1; + } + } + + if (lastIteration) { + this->_m_countIterations = 0; + } else { + ++this->_m_countIterations; + } + + return true; + } + +} /* namespace asr */ +} /* namespace audio */ +} /* namespace app */ +} /* namespace arm */
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/Wav2LetterPreprocess.cc b/source/use_case/kws_asr/src/Wav2LetterPreprocess.cc new file mode 100644 index 0000000..613ddb0 --- /dev/null +++ b/source/use_case/kws_asr/src/Wav2LetterPreprocess.cc @@ -0,0 +1,228 @@ +/* + * 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 "Wav2LetterPreprocess.hpp" + +#include "PlatformMath.hpp" +#include "TensorFlowLiteMicro.hpp" + +#include <algorithm> +#include <math.h> + +namespace arm { +namespace app { +namespace audio { +namespace asr { + + Preprocess::Preprocess( + const uint32_t numMfccFeatures, + const uint32_t windowLen, + const uint32_t windowStride, + const uint32_t numMfccVectors): + _m_mfcc(numMfccFeatures, windowLen), + _m_mfccBuf(numMfccFeatures, numMfccVectors), + _m_delta1Buf(numMfccFeatures, numMfccVectors), + _m_delta2Buf(numMfccFeatures, numMfccVectors), + _m_windowLen(windowLen), + _m_windowStride(windowStride), + _m_numMfccFeats(numMfccFeatures), + _m_numFeatVectors(numMfccVectors), + _m_window() + { + if (numMfccFeatures > 0 && windowLen > 0) { + this->_m_mfcc.Init(); + } + } + + bool Preprocess::Invoke( + const int16_t* audioData, + const uint32_t audioDataLen, + TfLiteTensor* tensor) + { + this->_m_window = SlidingWindow<const int16_t>( + audioData, audioDataLen, + this->_m_windowLen, this->_m_windowStride); + + uint32_t mfccBufIdx = 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 int16_t* mfccWindow = this->_m_window.Next(); + auto mfccAudioData = std::vector<int16_t>( + 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 adding MFCC for zeros. */ + if (mfccBufIdx != this->_m_numFeatVectors) { + std::vector<int16_t> zerosWindow = std::vector<int16_t>(this->_m_windowLen, 0); + std::vector<float> mfccZeros = this->_m_mfcc.MfccCompute(zerosWindow); + + while (mfccBufIdx != this->_m_numFeatVectors) { + memcpy(&this->_m_mfccBuf(0, mfccBufIdx), + mfccZeros.data(), sizeof(float) * _m_numMfccFeats); + ++mfccBufIdx; + } + } + + /* Compute first and second order deltas from MFCCs. */ + this->_ComputeDeltas(this->_m_mfccBuf, + this->_m_delta1Buf, + this->_m_delta2Buf); + + /* Normalise. */ + this->_Normalise(); + + /* Quantise. */ + QuantParams quantParams = GetTensorQuantParams(tensor); + + if (0 == quantParams.scale) { + printf_err("Quantisation scale can't be 0\n"); + return false; + } + + switch(tensor->type) { + case kTfLiteUInt8: + return this->_Quantise<uint8_t>( + tflite::GetTensorData<uint8_t>(tensor), tensor->bytes, + quantParams.scale, quantParams.offset); + case kTfLiteInt8: + return this->_Quantise<int8_t>( + tflite::GetTensorData<int8_t>(tensor), tensor->bytes, + quantParams.scale, quantParams.offset); + default: + printf_err("Unsupported tensor type %s\n", + TfLiteTypeGetName(tensor->type)); + } + + return false; + } + + 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 math::MathUtils::MeanF32(vec.begin(), vec.totalSize()); + } + + float Preprocess::_GetStdDev(Array2d<float>& vec, const float mean) + { + return math::MathUtils::StdDevF32(vec.begin(), vec.totalSize(), mean); + } + + void Preprocess::_NormaliseVec(Array2d<float>& vec) + { + auto mean = Preprocess::_GetMean(vec); + auto stddev = Preprocess::_GetStdDev(vec, mean); + + debug("Mean: %f, Stddev: %f\n", mean, stddev); + 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); + return std::min<float>(std::max<float>(val, minVal), maxVal); + } + +} /* namespace asr */ +} /* namespace audio */ +} /* namespace app */ +} /* namespace arm */
\ No newline at end of file |