diff options
author | George Gekov <george.gekov@arm.com> | 2021-08-16 11:32:10 +0100 |
---|---|---|
committer | Jim Flynn <jim.flynn@arm.com> | 2022-02-05 19:49:06 +0000 |
commit | 23c26277086c78704a17f0dae86da947816320c0 (patch) | |
tree | 88b02fd1fae3130256d059251788a7ef68d2831f /samples/SpeechRecognition/include/Wav2LetterMFCC.hpp | |
parent | 922b912fd2d462bac0809bac5669310ad1506310 (diff) | |
download | armnn-23c26277086c78704a17f0dae86da947816320c0.tar.gz |
MLECO-2079 Adding the C++ KWS example
Signed-off-by: Eanna O Cathain <eanna.ocathain@arm.com>
Change-Id: I81899bbfaada32f478c2e2fc6441eabb94d8d0fc
Diffstat (limited to 'samples/SpeechRecognition/include/Wav2LetterMFCC.hpp')
-rw-r--r-- | samples/SpeechRecognition/include/Wav2LetterMFCC.hpp | 78 |
1 files changed, 78 insertions, 0 deletions
diff --git a/samples/SpeechRecognition/include/Wav2LetterMFCC.hpp b/samples/SpeechRecognition/include/Wav2LetterMFCC.hpp new file mode 100644 index 0000000000..aa88aafb3b --- /dev/null +++ b/samples/SpeechRecognition/include/Wav2LetterMFCC.hpp @@ -0,0 +1,78 @@ +// +// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// +#pragma once + +#include "MFCC.hpp" + +/* Class to provide Wav2Letter specific MFCC calculation requirements. */ +class Wav2LetterMFCC : public MFCC +{ + +public: + explicit Wav2LetterMFCC(const MfccParams& params) + : MFCC(params) + {} + + Wav2LetterMFCC() = delete; + ~Wav2LetterMFCC() = default; + +protected: + + /** + * @brief Overrides base class implementation of this function. + * @param[in] fftVec Vector populated with FFT magnitudes + * @param[in] melFilterBank 2D Vector with filter bank weights + * @param[in] filterBankFilterFirst Vector containing the first indices of filter bank + * to be used for each bin. + * @param[in] filterBankFilterLast Vector containing the last indices of filter bank + * to be used for each bin. + * @param[out] melEnergies Pre-allocated vector of MEL energies to be + * populated. + * @return true if successful, false otherwise + */ + bool 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) override; + + /** + * @brief Override for the base class implementation convert mel + * energies to logarithmic scale. The difference from + * default behaviour is that the power is converted to dB + * and subsequently clamped. + * @param[in,out] melEnergies 1D vector of Mel energies + **/ + void ConvertToLogarithmicScale(std::vector<float>& melEnergies) override; + + /** + * @brief Create a matrix used to calculate Discrete Cosine + * Transform. Override for the base class' default + * implementation as the first and last elements + * use a different normaliser. + * @param[in] inputLength input length of the buffer on which + * DCT will be performed + * @param[in] coefficientCount Total coefficients per input length. + * @return 1D vector with inputLength x coefficientCount elements + * populated with DCT coefficients. + */ + std::vector<float> CreateDCTMatrix(int32_t inputLength, + int32_t coefficientCount) override; + + /** + * @brief Given the low and high Mel values, get the normaliser + * for weights to be applied when populating the filter + * bank. Override for the base class implementation. + * @param[in] leftMel Low Mel frequency value. + * @param[in] rightMel High Mel frequency value. + * @param[in] useHTKMethod bool to signal if HTK method is to be + * used for calculation. + * @return Value to use for normalising. + */ + float GetMelFilterBankNormaliser(const float& leftMel, + const float& rightMel, + bool useHTKMethod) override; +};
\ No newline at end of file |