Richard Burton | 0055346 | 2021-11-10 16:27:14 +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 "PlatformMath.hpp" |
| 18 | #include <cstdint> |
| 19 | #include <vector> |
| 20 | #include <array> |
| 21 | #include <tuple> |
| 22 | |
| 23 | namespace arm { |
| 24 | namespace app { |
| 25 | namespace rnn { |
| 26 | |
| 27 | using vec1D32F = std::vector<float>; |
| 28 | using vec2D32F = std::vector<vec1D32F>; |
| 29 | using arrHp = std::array<float, 2>; |
| 30 | using math::FftInstance; |
| 31 | using math::FftType; |
| 32 | |
| 33 | class FrameFeatures { |
| 34 | public: |
| 35 | bool m_silence{false}; /* If frame contains silence or not. */ |
| 36 | vec1D32F m_featuresVec{}; /* Calculated feature vector to feed to model. */ |
| 37 | vec1D32F m_fftX{}; /* Vector of floats arranged to represent complex numbers. */ |
| 38 | vec1D32F m_fftP{}; /* Vector of floats arranged to represent complex numbers. */ |
| 39 | vec1D32F m_Ex{}; /* Spectral band energy for audio x. */ |
| 40 | vec1D32F m_Ep{}; /* Spectral band energy for pitch p. */ |
| 41 | vec1D32F m_Exp{}; /* Correlated spectral energy between x and p. */ |
| 42 | }; |
| 43 | |
| 44 | /** |
| 45 | * @brief RNNoise pre and post processing class based on the 2018 paper from |
| 46 | * Jan-Marc Valin. Recommended reading: |
| 47 | * - https://jmvalin.ca/demo/rnnoise/ |
| 48 | * - https://arxiv.org/abs/1709.08243 |
| 49 | **/ |
| 50 | class RNNoiseProcess { |
| 51 | /* Public interface */ |
| 52 | public: |
| 53 | RNNoiseProcess(); |
| 54 | ~RNNoiseProcess() = default; |
| 55 | |
| 56 | /** |
| 57 | * @brief Calculates the features from a given audio buffer ready to be sent to RNNoise model. |
| 58 | * @param[in] audioData Pointer to the floating point vector |
| 59 | * with audio data (within the numerical |
| 60 | * limits of int16_t type). |
| 61 | * @param[in] audioLen Number of elements in the audio window. |
| 62 | * @param[out] features FrameFeatures object reference. |
| 63 | **/ |
| 64 | void PreprocessFrame(const float* audioData, |
| 65 | size_t audioLen, |
| 66 | FrameFeatures& features); |
| 67 | |
| 68 | /** |
| 69 | * @brief Use the RNNoise model output gain values with pre-processing features |
| 70 | * to generate audio with noise suppressed. |
| 71 | * @param[in] modelOutput Output gain values from model. |
| 72 | * @param[in] features Calculated features from pre-processing step. |
| 73 | * @param[out] outFrame Output frame to be populated. |
| 74 | **/ |
| 75 | void PostProcessFrame(vec1D32F& modelOutput, FrameFeatures& features, vec1D32F& outFrame); |
| 76 | |
| 77 | |
| 78 | /* Public constants */ |
| 79 | public: |
| 80 | static constexpr uint32_t FRAME_SIZE_SHIFT{2}; |
Richard Burton | 033c915 | 2021-12-07 14:04:44 +0000 | [diff] [blame^] | 81 | static constexpr uint32_t FRAME_SIZE{512}; |
Richard Burton | 0055346 | 2021-11-10 16:27:14 +0000 | [diff] [blame] | 82 | static constexpr uint32_t WINDOW_SIZE{2 * FRAME_SIZE}; |
| 83 | static constexpr uint32_t FREQ_SIZE{FRAME_SIZE + 1}; |
| 84 | |
Richard Burton | 033c915 | 2021-12-07 14:04:44 +0000 | [diff] [blame^] | 85 | static constexpr uint32_t PITCH_MIN_PERIOD{64}; |
| 86 | static constexpr uint32_t PITCH_MAX_PERIOD{820}; |
| 87 | static constexpr uint32_t PITCH_FRAME_SIZE{1024}; |
Richard Burton | 0055346 | 2021-11-10 16:27:14 +0000 | [diff] [blame] | 88 | static constexpr uint32_t PITCH_BUF_SIZE{PITCH_MAX_PERIOD + PITCH_FRAME_SIZE}; |
| 89 | |
| 90 | static constexpr uint32_t NB_BANDS{22}; |
| 91 | static constexpr uint32_t CEPS_MEM{8}; |
| 92 | static constexpr uint32_t NB_DELTA_CEPS{6}; |
| 93 | |
| 94 | static constexpr uint32_t NB_FEATURES{NB_BANDS + 3*NB_DELTA_CEPS + 2}; |
| 95 | |
| 96 | /* Private functions */ |
| 97 | private: |
| 98 | |
| 99 | /** |
| 100 | * @brief Initialises the half window and DCT tables. |
| 101 | */ |
| 102 | void InitTables(); |
| 103 | |
| 104 | /** |
| 105 | * @brief Applies a bi-quadratic filter over the audio window. |
| 106 | * @param[in] bHp Constant coefficient set b (arrHp type). |
| 107 | * @param[in] aHp Constant coefficient set a (arrHp type). |
| 108 | * @param[in,out] memHpX Coefficients populated by this function. |
| 109 | * @param[in,out] audioWindow Floating point vector with audio data. |
| 110 | **/ |
| 111 | void BiQuad( |
| 112 | const arrHp& bHp, |
| 113 | const arrHp& aHp, |
| 114 | arrHp& memHpX, |
| 115 | vec1D32F& audioWindow); |
| 116 | |
| 117 | /** |
| 118 | * @brief Computes features from the "filtered" audio window. |
| 119 | * @param[in] audioWindow Floating point vector with audio data. |
| 120 | * @param[out] features FrameFeatures object reference. |
| 121 | **/ |
| 122 | void ComputeFrameFeatures(vec1D32F& audioWindow, FrameFeatures& features); |
| 123 | |
| 124 | /** |
| 125 | * @brief Runs analysis on the audio buffer. |
| 126 | * @param[in] audioWindow Floating point vector with audio data. |
| 127 | * @param[out] fft Floating point FFT vector containing real and |
| 128 | * imaginary pairs of elements. NOTE: this vector |
| 129 | * does not contain the mirror image (conjugates) |
| 130 | * part of the spectrum. |
| 131 | * @param[out] energy Computed energy for each band in the Bark scale. |
| 132 | * @param[out] analysisMem Buffer sequentially, but partially, |
| 133 | * populated with new audio data. |
| 134 | **/ |
| 135 | void FrameAnalysis( |
| 136 | const vec1D32F& audioWindow, |
| 137 | vec1D32F& fft, |
| 138 | vec1D32F& energy, |
| 139 | vec1D32F& analysisMem); |
| 140 | |
| 141 | /** |
| 142 | * @brief Applies the window function, in-place, over the given |
| 143 | * floating point buffer. |
| 144 | * @param[in,out] x Buffer the window will be applied to. |
| 145 | **/ |
| 146 | void ApplyWindow(vec1D32F& x); |
| 147 | |
| 148 | /** |
| 149 | * @brief Computes the FFT for a given vector. |
| 150 | * @param[in] x Vector to compute the FFT from. |
| 151 | * @param[out] fft Floating point FFT vector containing real and |
| 152 | * imaginary pairs of elements. NOTE: this vector |
| 153 | * does not contain the mirror image (conjugates) |
| 154 | * part of the spectrum. |
| 155 | **/ |
| 156 | void ForwardTransform( |
| 157 | vec1D32F& x, |
| 158 | vec1D32F& fft); |
| 159 | |
| 160 | /** |
| 161 | * @brief Computes band energy for each of the 22 Bark scale bands. |
| 162 | * @param[in] fft_X FFT spectrum (as computed by ForwardTransform). |
| 163 | * @param[out] bandE Vector with 22 elements populated with energy for |
| 164 | * each band. |
| 165 | **/ |
| 166 | void ComputeBandEnergy(const vec1D32F& fft_X, vec1D32F& bandE); |
| 167 | |
| 168 | /** |
| 169 | * @brief Computes band energy correlation. |
| 170 | * @param[in] X FFT vector X. |
| 171 | * @param[in] P FFT vector P. |
| 172 | * @param[out] bandC Vector with 22 elements populated with band energy |
| 173 | * correlation for the two input FFT vectors. |
| 174 | **/ |
| 175 | void ComputeBandCorr(const vec1D32F& X, const vec1D32F& P, vec1D32F& bandC); |
| 176 | |
| 177 | /** |
| 178 | * @brief Performs pitch auto-correlation for a given vector for |
| 179 | * given lag. |
| 180 | * @param[in] x Input vector. |
| 181 | * @param[out] ac Auto-correlation output vector. |
| 182 | * @param[in] lag Lag value. |
| 183 | * @param[in] n Number of elements to consider for correlation |
| 184 | * computation. |
| 185 | **/ |
| 186 | void AutoCorr(const vec1D32F &x, |
| 187 | vec1D32F &ac, |
| 188 | size_t lag, |
| 189 | size_t n); |
| 190 | |
| 191 | /** |
| 192 | * @brief Computes pitch cross-correlation. |
| 193 | * @param[in] x Input vector 1. |
| 194 | * @param[in] y Input vector 2. |
George Gekov | a2b0fc2 | 2021-11-08 16:30:43 +0000 | [diff] [blame] | 195 | * @param[out] xCorr Cross-correlation output vector. |
Richard Burton | 0055346 | 2021-11-10 16:27:14 +0000 | [diff] [blame] | 196 | * @param[in] len Number of elements to consider for correlation. |
| 197 | * computation. |
| 198 | * @param[in] maxPitch Maximum pitch. |
| 199 | **/ |
| 200 | void PitchXCorr( |
| 201 | const vec1D32F& x, |
| 202 | const vec1D32F& y, |
George Gekov | a2b0fc2 | 2021-11-08 16:30:43 +0000 | [diff] [blame] | 203 | vec1D32F& xCorr, |
Richard Burton | 0055346 | 2021-11-10 16:27:14 +0000 | [diff] [blame] | 204 | size_t len, |
| 205 | size_t maxPitch); |
| 206 | |
| 207 | /** |
| 208 | * @brief Computes "Linear Predictor Coefficients". |
| 209 | * @param[in] ac Correlation vector. |
| 210 | * @param[in] p Number of elements of input vector to consider. |
| 211 | * @param[out] lpc Output coefficients vector. |
| 212 | **/ |
| 213 | void LPC(const vec1D32F& ac, int32_t p, vec1D32F& lpc); |
| 214 | |
| 215 | /** |
| 216 | * @brief Custom FIR implementation. |
| 217 | * @param[in] num FIR coefficient vector. |
| 218 | * @param[in] N Number of elements. |
| 219 | * @param[out] x Vector to be be processed. |
| 220 | **/ |
| 221 | void Fir5(const vec1D32F& num, uint32_t N, vec1D32F& x); |
| 222 | |
| 223 | /** |
| 224 | * @brief Down-sample the pitch buffer. |
| 225 | * @param[in,out] pitchBuf Pitch buffer. |
| 226 | * @param[in] pitchBufSz Buffer size. |
| 227 | **/ |
| 228 | void PitchDownsample(vec1D32F& pitchBuf, size_t pitchBufSz); |
| 229 | |
| 230 | /** |
| 231 | * @brief Pitch search function. |
| 232 | * @param[in] xLP Shifted pitch buffer input. |
| 233 | * @param[in] y Pitch buffer input. |
| 234 | * @param[in] len Length to search for. |
| 235 | * @param[in] maxPitch Maximum pitch. |
| 236 | * @return pitch index. |
| 237 | **/ |
| 238 | int PitchSearch(vec1D32F& xLp, vec1D32F& y, uint32_t len, uint32_t maxPitch); |
| 239 | |
| 240 | /** |
| 241 | * @brief Finds the "best" pitch from the buffer. |
| 242 | * @param[in] xCorr Pitch correlation vector. |
| 243 | * @param[in] y Pitch buffer input. |
| 244 | * @param[in] len Length to search for. |
| 245 | * @param[in] maxPitch Maximum pitch. |
| 246 | * @return pitch array (2 elements). |
| 247 | **/ |
| 248 | arrHp FindBestPitch(vec1D32F& xCorr, vec1D32F& y, uint32_t len, uint32_t maxPitch); |
| 249 | |
| 250 | /** |
| 251 | * @brief Remove pitch period doubling errors. |
| 252 | * @param[in,out] pitchBuf Pitch buffer vector. |
| 253 | * @param[in] maxPeriod Maximum period. |
| 254 | * @param[in] minPeriod Minimum period. |
| 255 | * @param[in] frameSize Frame size. |
| 256 | * @param[in] pitchIdx0_ Pitch index 0. |
| 257 | * @return pitch index. |
| 258 | **/ |
| 259 | int RemoveDoubling( |
| 260 | vec1D32F& pitchBuf, |
| 261 | uint32_t maxPeriod, |
| 262 | uint32_t minPeriod, |
| 263 | uint32_t frameSize, |
| 264 | size_t pitchIdx0_); |
| 265 | |
| 266 | /** |
| 267 | * @brief Computes pitch gain. |
| 268 | * @param[in] xy Single xy cross correlation value. |
| 269 | * @param[in] xx Single xx auto correlation value. |
| 270 | * @param[in] yy Single yy auto correlation value. |
| 271 | * @return Calculated pitch gain. |
| 272 | **/ |
| 273 | float ComputePitchGain(float xy, float xx, float yy); |
| 274 | |
| 275 | /** |
| 276 | * @brief Computes DCT vector from the given input. |
| 277 | * @param[in] input Input vector. |
| 278 | * @param[out] output Output vector with DCT coefficients. |
| 279 | **/ |
| 280 | void DCT(vec1D32F& input, vec1D32F& output); |
| 281 | |
| 282 | /** |
| 283 | * @brief Perform inverse fourier transform on complex spectral vector. |
| 284 | * @param[out] out Output vector. |
| 285 | * @param[in] fftXIn Vector of floats arranged to represent complex numbers interleaved. |
| 286 | **/ |
| 287 | void InverseTransform(vec1D32F& out, vec1D32F& fftXIn); |
| 288 | |
| 289 | /** |
| 290 | * @brief Perform pitch filtering. |
| 291 | * @param[in] features Object with pre-processing calculated frame features. |
| 292 | * @param[in] g Gain values. |
| 293 | **/ |
| 294 | void PitchFilter(FrameFeatures& features, vec1D32F& g); |
| 295 | |
| 296 | /** |
| 297 | * @brief Interpolate the band gain values. |
| 298 | * @param[out] g Gain values. |
| 299 | * @param[in] bandE Vector with 22 elements populated with energy for |
| 300 | * each band. |
| 301 | **/ |
| 302 | void InterpBandGain(vec1D32F& g, vec1D32F& bandE); |
| 303 | |
| 304 | /** |
| 305 | * @brief Create de-noised frame. |
| 306 | * @param[out] outFrame Output vector for storing the created audio frame. |
| 307 | * @param[in] fftY Gain adjusted complex spectral vector. |
| 308 | */ |
| 309 | void FrameSynthesis(vec1D32F& outFrame, vec1D32F& fftY); |
| 310 | |
| 311 | /* Private objects */ |
| 312 | private: |
| 313 | FftInstance m_fftInstReal; /* FFT instance for real numbers */ |
| 314 | FftInstance m_fftInstCmplx; /* FFT instance for complex numbers */ |
| 315 | vec1D32F m_halfWindow; /* Window coefficients */ |
| 316 | vec1D32F m_dctTable; /* DCT table */ |
| 317 | vec1D32F m_analysisMem; /* Buffer used for frame analysis */ |
| 318 | vec2D32F m_cepstralMem; /* Cepstral coefficients */ |
| 319 | size_t m_memId; /* memory ID */ |
| 320 | vec1D32F m_synthesisMem; /* Synthesis mem (used by post-processing) */ |
| 321 | vec1D32F m_pitchBuf; /* Pitch buffer */ |
| 322 | float m_lastGain; /* Last gain calculated */ |
| 323 | int m_lastPeriod; /* Last period calculated */ |
| 324 | arrHp m_memHpX; /* HpX coefficients. */ |
| 325 | vec1D32F m_lastGVec; /* Last gain vector (used by post-processing) */ |
| 326 | |
| 327 | /* Constants */ |
| 328 | const std::array <uint32_t, NB_BANDS> m_eband5ms { |
| 329 | 0, 1, 2, 3, 4, 5, 6, 7, 8, 10, 12, |
| 330 | 14, 16, 20, 24, 28, 34, 40, 48, 60, 78, 100}; |
| 331 | |
| 332 | }; |
| 333 | |
| 334 | |
| 335 | } /* namespace rnn */ |
| 336 | } /* namspace app */ |
| 337 | } /* namespace arm */ |