henrik.lundin@webrtc.org | 9a40081 | 2013-01-29 12:09:21 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. |
| 3 | * |
| 4 | * Use of this source code is governed by a BSD-style license |
| 5 | * that can be found in the LICENSE file in the root of the source |
| 6 | * tree. An additional intellectual property rights grant can be found |
| 7 | * in the file PATENTS. All contributing project authors may |
| 8 | * be found in the AUTHORS file in the root of the source tree. |
| 9 | */ |
| 10 | |
| 11 | #include "webrtc/modules/audio_coding/neteq4/background_noise.h" |
| 12 | |
| 13 | #include <assert.h> |
pbos@webrtc.org | 3f45c2e | 2013-08-05 16:22:53 +0000 | [diff] [blame] | 14 | #include <string.h> // memcpy |
henrik.lundin@webrtc.org | 9a40081 | 2013-01-29 12:09:21 +0000 | [diff] [blame] | 15 | |
| 16 | #include <algorithm> // min, max |
henrik.lundin@webrtc.org | 9a40081 | 2013-01-29 12:09:21 +0000 | [diff] [blame] | 17 | |
| 18 | #include "webrtc/common_audio/signal_processing/include/signal_processing_library.h" |
| 19 | #include "webrtc/modules/audio_coding/neteq4/audio_multi_vector.h" |
| 20 | #include "webrtc/modules/audio_coding/neteq4/post_decode_vad.h" |
| 21 | |
| 22 | namespace webrtc { |
| 23 | |
pbos@webrtc.org | dd1b19d | 2013-07-31 15:54:00 +0000 | [diff] [blame] | 24 | BackgroundNoise::BackgroundNoise(size_t num_channels) |
| 25 | : num_channels_(num_channels), |
| 26 | channel_parameters_(new ChannelParameters[num_channels_]), |
| 27 | mode_(kBgnOn) { |
| 28 | Reset(); |
| 29 | } |
| 30 | |
| 31 | BackgroundNoise::~BackgroundNoise() {} |
| 32 | |
henrik.lundin@webrtc.org | 9a40081 | 2013-01-29 12:09:21 +0000 | [diff] [blame] | 33 | void BackgroundNoise::Reset() { |
| 34 | initialized_ = false; |
| 35 | for (size_t channel = 0; channel < num_channels_; ++channel) { |
| 36 | channel_parameters_[channel].Reset(); |
| 37 | } |
| 38 | // Keep _bgnMode as it is. |
| 39 | } |
| 40 | |
henrik.lundin@webrtc.org | 0e9c399 | 2013-09-30 20:38:44 +0000 | [diff] [blame^] | 41 | void BackgroundNoise::Update(const AudioMultiVector& input, |
henrik.lundin@webrtc.org | 9a40081 | 2013-01-29 12:09:21 +0000 | [diff] [blame] | 42 | const PostDecodeVad& vad) { |
| 43 | if (vad.running() && vad.active_speech()) { |
| 44 | // Do not update the background noise parameters if we know that the signal |
| 45 | // is active speech. |
| 46 | return; |
| 47 | } |
| 48 | |
| 49 | int32_t auto_correlation[kMaxLpcOrder + 1]; |
| 50 | int16_t fiter_output[kMaxLpcOrder + kResidualLength]; |
| 51 | int16_t reflection_coefficients[kMaxLpcOrder]; |
| 52 | int16_t lpc_coefficients[kMaxLpcOrder + 1]; |
| 53 | |
| 54 | for (size_t channel_ix = 0; channel_ix < num_channels_; ++channel_ix) { |
| 55 | ChannelParameters& parameters = channel_parameters_[channel_ix]; |
| 56 | int16_t temp_signal_array[kVecLen + kMaxLpcOrder] = {0}; |
| 57 | int16_t* temp_signal = &temp_signal_array[kMaxLpcOrder]; |
| 58 | memcpy(temp_signal, |
| 59 | &input[channel_ix][input.Size() - kVecLen], |
| 60 | sizeof(int16_t) * kVecLen); |
| 61 | |
| 62 | int32_t sample_energy = CalculateAutoCorrelation(temp_signal, kVecLen, |
| 63 | auto_correlation); |
| 64 | |
| 65 | if ((!vad.running() && |
| 66 | sample_energy < parameters.energy_update_threshold) || |
| 67 | (vad.running() && !vad.active_speech())) { |
| 68 | // Generate LPC coefficients. |
| 69 | if (auto_correlation[0] > 0) { |
| 70 | // Regardless of whether the filter is actually updated or not, |
| 71 | // update energy threshold levels, since we have in fact observed |
| 72 | // a low energy signal. |
| 73 | if (sample_energy < parameters.energy_update_threshold) { |
| 74 | // Never go under 1.0 in average sample energy. |
| 75 | parameters.energy_update_threshold = std::max(sample_energy, 1); |
| 76 | parameters.low_energy_update_threshold = 0; |
| 77 | } |
| 78 | |
| 79 | // Only update BGN if filter is stable, i.e., if return value from |
| 80 | // Levinson-Durbin function is 1. |
| 81 | if (WebRtcSpl_LevinsonDurbin(auto_correlation, lpc_coefficients, |
| 82 | reflection_coefficients, |
| 83 | kMaxLpcOrder) != 1) { |
| 84 | return; |
| 85 | } |
| 86 | } else { |
| 87 | // Center value in auto-correlation is not positive. Do not update. |
| 88 | return; |
| 89 | } |
| 90 | |
| 91 | // Generate the CNG gain factor by looking at the energy of the residual. |
| 92 | WebRtcSpl_FilterMAFastQ12(temp_signal + kVecLen - kResidualLength, |
| 93 | fiter_output, lpc_coefficients, |
| 94 | kMaxLpcOrder + 1, kResidualLength); |
| 95 | int32_t residual_energy = WebRtcSpl_DotProductWithScale(fiter_output, |
| 96 | fiter_output, |
| 97 | kResidualLength, |
| 98 | 0); |
| 99 | |
| 100 | // Check spectral flatness. |
| 101 | // Comparing the residual variance with the input signal variance tells |
| 102 | // if the spectrum is flat or not. |
| 103 | // If 20 * residual_energy >= sample_energy << 6, the spectrum is flat |
| 104 | // enough. Also ensure that the energy is non-zero. |
| 105 | if ((residual_energy * 20 >= (sample_energy << 6)) && |
| 106 | (sample_energy > 0)) { |
| 107 | // Spectrum is flat enough; save filter parameters. |
| 108 | // |temp_signal| + |kVecLen| - |kMaxLpcOrder| points at the first of the |
| 109 | // |kMaxLpcOrder| samples in the residual signal, which will form the |
| 110 | // filter state for the next noise generation. |
| 111 | SaveParameters(channel_ix, lpc_coefficients, |
| 112 | temp_signal + kVecLen - kMaxLpcOrder, sample_energy, |
| 113 | residual_energy); |
| 114 | } |
| 115 | } else { |
| 116 | // Will only happen if post-decode VAD is disabled and |sample_energy| is |
| 117 | // not low enough. Increase the threshold for update so that it increases |
| 118 | // by a factor 4 in 4 seconds. |
| 119 | IncrementEnergyThreshold(channel_ix, sample_energy); |
| 120 | } |
| 121 | } |
| 122 | return; |
| 123 | } |
| 124 | |
| 125 | int32_t BackgroundNoise::Energy(size_t channel) const { |
| 126 | assert(channel < num_channels_); |
| 127 | return channel_parameters_[channel].energy; |
| 128 | } |
| 129 | |
| 130 | void BackgroundNoise::SetMuteFactor(size_t channel, int16_t value) { |
| 131 | assert(channel < num_channels_); |
| 132 | channel_parameters_[channel].mute_factor = value; |
| 133 | } |
| 134 | |
| 135 | int16_t BackgroundNoise::MuteFactor(size_t channel) const { |
| 136 | assert(channel < num_channels_); |
| 137 | return channel_parameters_[channel].mute_factor; |
| 138 | } |
| 139 | |
| 140 | const int16_t* BackgroundNoise::Filter(size_t channel) const { |
| 141 | assert(channel < num_channels_); |
| 142 | return channel_parameters_[channel].filter; |
| 143 | } |
| 144 | |
| 145 | const int16_t* BackgroundNoise::FilterState(size_t channel) const { |
| 146 | assert(channel < num_channels_); |
| 147 | return channel_parameters_[channel].filter_state; |
| 148 | } |
| 149 | |
| 150 | void BackgroundNoise::SetFilterState(size_t channel, const int16_t* input, |
| 151 | size_t length) { |
| 152 | assert(channel < num_channels_); |
| 153 | length = std::min(length, static_cast<size_t>(kMaxLpcOrder)); |
| 154 | memcpy(channel_parameters_[channel].filter_state, input, |
| 155 | length * sizeof(int16_t)); |
| 156 | } |
| 157 | |
| 158 | int16_t BackgroundNoise::Scale(size_t channel) const { |
| 159 | assert(channel < num_channels_); |
| 160 | return channel_parameters_[channel].scale; |
| 161 | } |
| 162 | int16_t BackgroundNoise::ScaleShift(size_t channel) const { |
| 163 | assert(channel < num_channels_); |
| 164 | return channel_parameters_[channel].scale_shift; |
| 165 | } |
| 166 | |
| 167 | int32_t BackgroundNoise::CalculateAutoCorrelation( |
turaj@webrtc.org | 045e45e | 2013-09-20 16:25:28 +0000 | [diff] [blame] | 168 | const int16_t* signal, int length, int32_t* auto_correlation) const { |
henrik.lundin@webrtc.org | 9a40081 | 2013-01-29 12:09:21 +0000 | [diff] [blame] | 169 | int16_t signal_max = WebRtcSpl_MaxAbsValueW16(signal, length); |
| 170 | int correlation_scale = kLogVecLen - |
| 171 | WebRtcSpl_NormW32(signal_max * signal_max); |
| 172 | correlation_scale = std::max(0, correlation_scale); |
| 173 | |
| 174 | static const int kCorrelationStep = -1; |
turaj@webrtc.org | 045e45e | 2013-09-20 16:25:28 +0000 | [diff] [blame] | 175 | WebRtcSpl_CrossCorrelation(auto_correlation, signal, signal, length, |
| 176 | kMaxLpcOrder + 1, correlation_scale, |
henrik.lundin@webrtc.org | 9a40081 | 2013-01-29 12:09:21 +0000 | [diff] [blame] | 177 | kCorrelationStep); |
| 178 | |
| 179 | // Number of shifts to normalize energy to energy/sample. |
| 180 | int energy_sample_shift = kLogVecLen - correlation_scale; |
| 181 | return auto_correlation[0] >> energy_sample_shift; |
| 182 | } |
| 183 | |
| 184 | void BackgroundNoise::IncrementEnergyThreshold(size_t channel, |
| 185 | int32_t sample_energy) { |
| 186 | // TODO(hlundin): Simplify the below threshold update. What this code |
| 187 | // does is simply "threshold += (increment * threshold) >> 16", but due |
| 188 | // to the limited-width operations, it is not exactly the same. The |
| 189 | // difference should be inaudible, but bit-exactness would not be |
| 190 | // maintained. |
| 191 | assert(channel < num_channels_); |
| 192 | ChannelParameters& parameters = channel_parameters_[channel]; |
| 193 | int32_t temp_energy = |
| 194 | WEBRTC_SPL_MUL_16_16_RSFT(kThresholdIncrement, |
| 195 | parameters.low_energy_update_threshold, 16); |
| 196 | temp_energy += kThresholdIncrement * |
| 197 | (parameters.energy_update_threshold & 0xFF); |
| 198 | temp_energy += (kThresholdIncrement * |
| 199 | ((parameters.energy_update_threshold>>8) & 0xFF)) << 8; |
| 200 | parameters.low_energy_update_threshold += temp_energy; |
| 201 | |
| 202 | parameters.energy_update_threshold += kThresholdIncrement * |
| 203 | (parameters.energy_update_threshold>>16); |
| 204 | parameters.energy_update_threshold += |
| 205 | parameters.low_energy_update_threshold >> 16; |
| 206 | parameters.low_energy_update_threshold = |
| 207 | parameters.low_energy_update_threshold & 0x0FFFF; |
| 208 | |
| 209 | // Update maximum energy. |
| 210 | // Decrease by a factor 1/1024 each time. |
| 211 | parameters.max_energy = parameters.max_energy - |
| 212 | (parameters.max_energy >> 10); |
| 213 | if (sample_energy > parameters.max_energy) { |
| 214 | parameters.max_energy = sample_energy; |
| 215 | } |
| 216 | |
| 217 | // Set |energy_update_threshold| to no less than 60 dB lower than |
| 218 | // |max_energy_|. Adding 524288 assures proper rounding. |
| 219 | int32_t energy_update_threshold = (parameters.max_energy + 524288) >> 20; |
| 220 | if (energy_update_threshold > parameters.energy_update_threshold) { |
| 221 | parameters.energy_update_threshold = energy_update_threshold; |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | void BackgroundNoise::SaveParameters(size_t channel, |
| 226 | const int16_t* lpc_coefficients, |
| 227 | const int16_t* filter_state, |
| 228 | int32_t sample_energy, |
| 229 | int32_t residual_energy) { |
| 230 | assert(channel < num_channels_); |
| 231 | ChannelParameters& parameters = channel_parameters_[channel]; |
| 232 | memcpy(parameters.filter, lpc_coefficients, |
| 233 | (kMaxLpcOrder+1) * sizeof(int16_t)); |
| 234 | memcpy(parameters.filter_state, filter_state, |
| 235 | kMaxLpcOrder * sizeof(int16_t)); |
| 236 | // Save energy level and update energy threshold levels. |
| 237 | // Never get under 1.0 in average sample energy. |
| 238 | parameters.energy = std::max(sample_energy, 1); |
| 239 | parameters.energy_update_threshold = parameters.energy; |
| 240 | parameters.low_energy_update_threshold = 0; |
| 241 | |
| 242 | // Normalize residual_energy to 29 or 30 bits before sqrt. |
| 243 | int norm_shift = WebRtcSpl_NormW32(residual_energy) - 1; |
| 244 | if (norm_shift & 0x1) { |
| 245 | norm_shift -= 1; // Even number of shifts required. |
| 246 | } |
| 247 | assert(norm_shift >= 0); // Should always be positive. |
| 248 | residual_energy = residual_energy << norm_shift; |
| 249 | |
| 250 | // Calculate scale and shift factor. |
| 251 | parameters.scale = WebRtcSpl_SqrtFloor(residual_energy); |
| 252 | // Add 13 to the |scale_shift_|, since the random numbers table is in |
| 253 | // Q13. |
| 254 | // TODO(hlundin): Move the "13" to where the |scale_shift_| is used? |
| 255 | parameters.scale_shift = 13 + ((kLogResidualLength + norm_shift) / 2); |
| 256 | |
| 257 | initialized_ = true; |
| 258 | } |
| 259 | |
| 260 | } // namespace webrtc |