2 * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved.
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.
11 #include "webrtc/modules/audio_coding/neteq/expand.h"
14 #include <string.h> // memset
16 #include <algorithm> // min, max
17 #include <limits> // numeric_limits<T>
19 #include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
20 #include "webrtc/modules/audio_coding/neteq/background_noise.h"
21 #include "webrtc/modules/audio_coding/neteq/dsp_helper.h"
22 #include "webrtc/modules/audio_coding/neteq/random_vector.h"
23 #include "webrtc/modules/audio_coding/neteq/sync_buffer.h"
27 void Expand::Reset() {
29 consecutive_expands_ = 0;
31 for (size_t ix = 0; ix < num_channels_; ++ix) {
32 channel_parameters_[ix].expand_vector0.Clear();
33 channel_parameters_[ix].expand_vector1.Clear();
37 int Expand::Process(AudioMultiVector* output) {
38 int16_t random_vector[kMaxSampleRate / 8000 * 120 + 30];
39 int16_t scaled_random_vector[kMaxSampleRate / 8000 * 125];
40 static const int kTempDataSize = 3600;
41 int16_t temp_data[kTempDataSize]; // TODO(hlundin) Remove this.
42 int16_t* voiced_vector_storage = temp_data;
43 int16_t* voiced_vector = &voiced_vector_storage[overlap_length_];
44 static const int kNoiseLpcOrder = BackgroundNoise::kMaxLpcOrder;
45 int16_t unvoiced_array_memory[kNoiseLpcOrder + kMaxSampleRate / 8000 * 125];
46 int16_t* unvoiced_vector = unvoiced_array_memory + kUnvoicedLpcOrder;
47 int16_t* noise_vector = unvoiced_array_memory + kNoiseLpcOrder;
49 int fs_mult = fs_hz_ / 8000;
52 // Perform initial setup if this is the first expansion since last reset.
53 AnalyzeSignal(random_vector);
54 first_expand_ = false;
56 // This is not the first expansion, parameters are already estimated.
57 // Extract a noise segment.
58 int16_t rand_length = max_lag_;
59 // This only applies to SWB where length could be larger than 256.
60 assert(rand_length <= kMaxSampleRate / 8000 * 120 + 30);
61 GenerateRandomVector(2, rand_length, random_vector);
69 // Generate a weighted vector with the current lag.
70 size_t expansion_vector_length = max_lag_ + overlap_length_;
71 size_t current_lag = expand_lags_[current_lag_index_];
72 // Copy lag+overlap data.
73 size_t expansion_vector_position = expansion_vector_length - current_lag -
75 size_t temp_length = current_lag + overlap_length_;
76 for (size_t channel_ix = 0; channel_ix < num_channels_; ++channel_ix) {
77 ChannelParameters& parameters = channel_parameters_[channel_ix];
78 if (current_lag_index_ == 0) {
79 // Use only expand_vector0.
80 assert(expansion_vector_position + temp_length <=
81 parameters.expand_vector0.Size());
82 memcpy(voiced_vector_storage,
83 ¶meters.expand_vector0[expansion_vector_position],
84 sizeof(int16_t) * temp_length);
85 } else if (current_lag_index_ == 1) {
86 // Mix 3/4 of expand_vector0 with 1/4 of expand_vector1.
87 WebRtcSpl_ScaleAndAddVectorsWithRound(
88 ¶meters.expand_vector0[expansion_vector_position], 3,
89 ¶meters.expand_vector1[expansion_vector_position], 1, 2,
90 voiced_vector_storage, static_cast<int>(temp_length));
91 } else if (current_lag_index_ == 2) {
92 // Mix 1/2 of expand_vector0 with 1/2 of expand_vector1.
93 assert(expansion_vector_position + temp_length <=
94 parameters.expand_vector0.Size());
95 assert(expansion_vector_position + temp_length <=
96 parameters.expand_vector1.Size());
97 WebRtcSpl_ScaleAndAddVectorsWithRound(
98 ¶meters.expand_vector0[expansion_vector_position], 1,
99 ¶meters.expand_vector1[expansion_vector_position], 1, 1,
100 voiced_vector_storage, static_cast<int>(temp_length));
103 // Get tapering window parameters. Values are in Q15.
104 int16_t muting_window, muting_window_increment;
105 int16_t unmuting_window, unmuting_window_increment;
106 if (fs_hz_ == 8000) {
107 muting_window = DspHelper::kMuteFactorStart8kHz;
108 muting_window_increment = DspHelper::kMuteFactorIncrement8kHz;
109 unmuting_window = DspHelper::kUnmuteFactorStart8kHz;
110 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement8kHz;
111 } else if (fs_hz_ == 16000) {
112 muting_window = DspHelper::kMuteFactorStart16kHz;
113 muting_window_increment = DspHelper::kMuteFactorIncrement16kHz;
114 unmuting_window = DspHelper::kUnmuteFactorStart16kHz;
115 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement16kHz;
116 } else if (fs_hz_ == 32000) {
117 muting_window = DspHelper::kMuteFactorStart32kHz;
118 muting_window_increment = DspHelper::kMuteFactorIncrement32kHz;
119 unmuting_window = DspHelper::kUnmuteFactorStart32kHz;
120 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement32kHz;
121 } else { // fs_ == 48000
122 muting_window = DspHelper::kMuteFactorStart48kHz;
123 muting_window_increment = DspHelper::kMuteFactorIncrement48kHz;
124 unmuting_window = DspHelper::kUnmuteFactorStart48kHz;
125 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement48kHz;
128 // Smooth the expanded if it has not been muted to a low amplitude and
129 // |current_voice_mix_factor| is larger than 0.5.
130 if ((parameters.mute_factor > 819) &&
131 (parameters.current_voice_mix_factor > 8192)) {
132 size_t start_ix = sync_buffer_->Size() - overlap_length_;
133 for (size_t i = 0; i < overlap_length_; i++) {
134 // Do overlap add between new vector and overlap.
135 (*sync_buffer_)[channel_ix][start_ix + i] =
136 (((*sync_buffer_)[channel_ix][start_ix + i] * muting_window) +
137 (((parameters.mute_factor * voiced_vector_storage[i]) >> 14) *
138 unmuting_window) + 16384) >> 15;
139 muting_window += muting_window_increment;
140 unmuting_window += unmuting_window_increment;
142 } else if (parameters.mute_factor == 0) {
143 // The expanded signal will consist of only comfort noise if
144 // mute_factor = 0. Set the output length to 15 ms for best noise
146 // TODO(hlundin): This has been disabled since the length of
147 // parameters.expand_vector0 and parameters.expand_vector1 no longer
148 // match with expand_lags_, causing invalid reads and writes. Is it a good
149 // idea to enable this again, and solve the vector size problem?
150 // max_lag_ = fs_mult * 120;
151 // expand_lags_[0] = fs_mult * 120;
152 // expand_lags_[1] = fs_mult * 120;
153 // expand_lags_[2] = fs_mult * 120;
157 // Filter |scaled_random_vector| through |ar_filter_|.
158 memcpy(unvoiced_vector - kUnvoicedLpcOrder, parameters.ar_filter_state,
159 sizeof(int16_t) * kUnvoicedLpcOrder);
160 int32_t add_constant = 0;
161 if (parameters.ar_gain_scale > 0) {
162 add_constant = 1 << (parameters.ar_gain_scale - 1);
164 WebRtcSpl_AffineTransformVector(scaled_random_vector, random_vector,
165 parameters.ar_gain, add_constant,
166 parameters.ar_gain_scale,
167 static_cast<int>(current_lag));
168 WebRtcSpl_FilterARFastQ12(scaled_random_vector, unvoiced_vector,
169 parameters.ar_filter, kUnvoicedLpcOrder + 1,
170 static_cast<int>(current_lag));
171 memcpy(parameters.ar_filter_state,
172 &(unvoiced_vector[current_lag - kUnvoicedLpcOrder]),
173 sizeof(int16_t) * kUnvoicedLpcOrder);
175 // Combine voiced and unvoiced contributions.
177 // Set a suitable cross-fading slope.
179 // <= 31 * fs_mult => go from 1 to 0 in about 8 ms;
180 // (>= 31 .. <= 63) * fs_mult => go from 1 to 0 in about 16 ms;
181 // >= 64 * fs_mult => go from 1 to 0 in about 32 ms.
182 // temp_shift = getbits(max_lag_) - 5.
183 int temp_shift = (31 - WebRtcSpl_NormW32(max_lag_)) - 5;
184 int16_t mix_factor_increment = 256 >> temp_shift;
186 mix_factor_increment = 0;
189 // Create combined signal by shifting in more and more of unvoiced part.
190 temp_shift = 8 - temp_shift; // = getbits(mix_factor_increment).
191 size_t temp_lenght = (parameters.current_voice_mix_factor -
192 parameters.voice_mix_factor) >> temp_shift;
193 temp_lenght = std::min(temp_lenght, current_lag);
194 DspHelper::CrossFade(voiced_vector, unvoiced_vector, temp_lenght,
195 ¶meters.current_voice_mix_factor,
196 mix_factor_increment, temp_data);
198 // End of cross-fading period was reached before end of expanded signal
199 // path. Mix the rest with a fixed mixing factor.
200 if (temp_lenght < current_lag) {
201 if (mix_factor_increment != 0) {
202 parameters.current_voice_mix_factor = parameters.voice_mix_factor;
204 int temp_scale = 16384 - parameters.current_voice_mix_factor;
205 WebRtcSpl_ScaleAndAddVectorsWithRound(
206 voiced_vector + temp_lenght, parameters.current_voice_mix_factor,
207 unvoiced_vector + temp_lenght, temp_scale, 14,
208 temp_data + temp_lenght, static_cast<int>(current_lag - temp_lenght));
211 // Select muting slope depending on how many consecutive expands we have
213 if (consecutive_expands_ == 3) {
214 // Let the mute factor decrease from 1.0 to 0.95 in 6.25 ms.
215 // mute_slope = 0.0010 / fs_mult in Q20.
216 parameters.mute_slope = std::max(parameters.mute_slope,
217 static_cast<int16_t>(1049 / fs_mult));
219 if (consecutive_expands_ == 7) {
220 // Let the mute factor decrease from 1.0 to 0.90 in 6.25 ms.
221 // mute_slope = 0.0020 / fs_mult in Q20.
222 parameters.mute_slope = std::max(parameters.mute_slope,
223 static_cast<int16_t>(2097 / fs_mult));
226 // Mute segment according to slope value.
227 if ((consecutive_expands_ != 0) || !parameters.onset) {
228 // Mute to the previous level, then continue with the muting.
229 WebRtcSpl_AffineTransformVector(temp_data, temp_data,
230 parameters.mute_factor, 8192,
231 14, static_cast<int>(current_lag));
234 DspHelper::MuteSignal(temp_data, parameters.mute_slope, current_lag);
236 // Shift by 6 to go from Q20 to Q14.
237 // TODO(hlundin): Adding 8192 before shifting 6 steps seems wrong.
239 int16_t gain = static_cast<int16_t>(16384 -
240 (((current_lag * parameters.mute_slope) + 8192) >> 6));
241 gain = ((gain * parameters.mute_factor) + 8192) >> 14;
243 // Guard against getting stuck with very small (but sometimes audible)
245 if ((consecutive_expands_ > 3) && (gain >= parameters.mute_factor)) {
246 parameters.mute_factor = 0;
248 parameters.mute_factor = gain;
253 // Background noise part.
254 GenerateBackgroundNoise(random_vector,
256 channel_parameters_[channel_ix].mute_slope,
259 unvoiced_array_memory);
261 // Add background noise to the combined voiced-unvoiced signal.
262 for (size_t i = 0; i < current_lag; i++) {
263 temp_data[i] = temp_data[i] + noise_vector[i];
265 if (channel_ix == 0) {
266 output->AssertSize(current_lag);
268 assert(output->Size() == current_lag);
270 memcpy(&(*output)[channel_ix][0], temp_data,
271 sizeof(temp_data[0]) * current_lag);
274 // Increase call number and cap it.
275 consecutive_expands_ = consecutive_expands_ >= kMaxConsecutiveExpands ?
276 kMaxConsecutiveExpands : consecutive_expands_ + 1;
280 void Expand::SetParametersForNormalAfterExpand() {
281 current_lag_index_ = 0;
282 lag_index_direction_ = 0;
283 stop_muting_ = true; // Do not mute signal any more.
286 void Expand::SetParametersForMergeAfterExpand() {
287 current_lag_index_ = -1; /* out of the 3 possible ones */
288 lag_index_direction_ = 1; /* make sure we get the "optimal" lag */
292 void Expand::InitializeForAnExpandPeriod() {
293 lag_index_direction_ = 1;
294 current_lag_index_ = -1;
295 stop_muting_ = false;
296 random_vector_->set_seed_increment(1);
297 consecutive_expands_ = 0;
298 for (size_t ix = 0; ix < num_channels_; ++ix) {
299 channel_parameters_[ix].current_voice_mix_factor = 16384; // 1.0 in Q14.
300 channel_parameters_[ix].mute_factor = 16384; // 1.0 in Q14.
301 // Start with 0 gain for background noise.
302 background_noise_->SetMuteFactor(ix, 0);
306 bool Expand::TooManyExpands() {
307 return consecutive_expands_ >= kMaxConsecutiveExpands;
310 void Expand::AnalyzeSignal(int16_t* random_vector) {
311 int32_t auto_correlation[kUnvoicedLpcOrder + 1];
312 int16_t reflection_coeff[kUnvoicedLpcOrder];
313 int16_t correlation_vector[kMaxSampleRate / 8000 * 102];
314 int best_correlation_index[kNumCorrelationCandidates];
315 int16_t best_correlation[kNumCorrelationCandidates];
316 int16_t best_distortion_index[kNumCorrelationCandidates];
317 int16_t best_distortion[kNumCorrelationCandidates];
318 int32_t correlation_vector2[(99 * kMaxSampleRate / 8000) + 1];
319 int32_t best_distortion_w32[kNumCorrelationCandidates];
320 static const int kNoiseLpcOrder = BackgroundNoise::kMaxLpcOrder;
321 int16_t unvoiced_array_memory[kNoiseLpcOrder + kMaxSampleRate / 8000 * 125];
322 int16_t* unvoiced_vector = unvoiced_array_memory + kUnvoicedLpcOrder;
324 int fs_mult = fs_hz_ / 8000;
326 // Pre-calculate common multiplications with fs_mult.
327 int fs_mult_4 = fs_mult * 4;
328 int fs_mult_20 = fs_mult * 20;
329 int fs_mult_120 = fs_mult * 120;
330 int fs_mult_dist_len = fs_mult * kDistortionLength;
331 int fs_mult_lpc_analysis_len = fs_mult * kLpcAnalysisLength;
333 const size_t signal_length = 256 * fs_mult;
334 const int16_t* audio_history =
335 &(*sync_buffer_)[0][sync_buffer_->Size() - signal_length];
338 InitializeForAnExpandPeriod();
340 // Calculate correlation in downsampled domain (4 kHz sample rate).
341 int16_t correlation_scale;
342 int correlation_length = 51; // TODO(hlundin): Legacy bit-exactness.
343 // If it is decided to break bit-exactness |correlation_length| should be
344 // initialized to the return value of Correlation().
345 Correlation(audio_history, signal_length, correlation_vector,
348 // Find peaks in correlation vector.
349 DspHelper::PeakDetection(correlation_vector, correlation_length,
350 kNumCorrelationCandidates, fs_mult,
351 best_correlation_index, best_correlation);
353 // Adjust peak locations; cross-correlation lags start at 2.5 ms
354 // (20 * fs_mult samples).
355 best_correlation_index[0] += fs_mult_20;
356 best_correlation_index[1] += fs_mult_20;
357 best_correlation_index[2] += fs_mult_20;
359 // Calculate distortion around the |kNumCorrelationCandidates| best lags.
360 int distortion_scale = 0;
361 for (int i = 0; i < kNumCorrelationCandidates; i++) {
362 int16_t min_index = std::max(fs_mult_20,
363 best_correlation_index[i] - fs_mult_4);
364 int16_t max_index = std::min(fs_mult_120 - 1,
365 best_correlation_index[i] + fs_mult_4);
366 best_distortion_index[i] = DspHelper::MinDistortion(
367 &(audio_history[signal_length - fs_mult_dist_len]), min_index,
368 max_index, fs_mult_dist_len, &best_distortion_w32[i]);
369 distortion_scale = std::max(16 - WebRtcSpl_NormW32(best_distortion_w32[i]),
372 // Shift the distortion values to fit in 16 bits.
373 WebRtcSpl_VectorBitShiftW32ToW16(best_distortion, kNumCorrelationCandidates,
374 best_distortion_w32, distortion_scale);
376 // Find the maximizing index |i| of the cost function
377 // f[i] = best_correlation[i] / best_distortion[i].
378 int32_t best_ratio = std::numeric_limits<int32_t>::min();
380 for (int i = 0; i < kNumCorrelationCandidates; ++i) {
382 if (best_distortion[i] > 0) {
383 ratio = (best_correlation[i] << 16) / best_distortion[i];
384 } else if (best_correlation[i] == 0) {
385 ratio = 0; // No correlation set result to zero.
387 ratio = std::numeric_limits<int32_t>::max(); // Denominator is zero.
389 if (ratio > best_ratio) {
395 int distortion_lag = best_distortion_index[best_index];
396 int correlation_lag = best_correlation_index[best_index];
397 max_lag_ = std::max(distortion_lag, correlation_lag);
399 // Calculate the exact best correlation in the range between
400 // |correlation_lag| and |distortion_lag|.
401 correlation_length = distortion_lag + 10;
402 correlation_length = std::min(correlation_length, fs_mult_120);
403 correlation_length = std::max(correlation_length, 60 * fs_mult);
405 int start_index = std::min(distortion_lag, correlation_lag);
406 int correlation_lags = WEBRTC_SPL_ABS_W16((distortion_lag-correlation_lag))
408 assert(correlation_lags <= 99 * fs_mult + 1); // Cannot be larger.
410 for (size_t channel_ix = 0; channel_ix < num_channels_; ++channel_ix) {
411 ChannelParameters& parameters = channel_parameters_[channel_ix];
412 // Calculate suitable scaling.
413 int16_t signal_max = WebRtcSpl_MaxAbsValueW16(
414 &audio_history[signal_length - correlation_length - start_index
416 correlation_length + start_index + correlation_lags - 1);
417 correlation_scale = ((31 - WebRtcSpl_NormW32(signal_max * signal_max))
418 + (31 - WebRtcSpl_NormW32(correlation_length))) - 31;
419 correlation_scale = std::max(static_cast<int16_t>(0), correlation_scale);
421 // Calculate the correlation, store in |correlation_vector2|.
422 WebRtcSpl_CrossCorrelation(
424 &(audio_history[signal_length - correlation_length]),
425 &(audio_history[signal_length - correlation_length - start_index]),
426 correlation_length, correlation_lags, correlation_scale, -1);
428 // Find maximizing index.
429 best_index = WebRtcSpl_MaxIndexW32(correlation_vector2, correlation_lags);
430 int32_t max_correlation = correlation_vector2[best_index];
431 // Compensate index with start offset.
432 best_index = best_index + start_index;
434 // Calculate energies.
435 int32_t energy1 = WebRtcSpl_DotProductWithScale(
436 &(audio_history[signal_length - correlation_length]),
437 &(audio_history[signal_length - correlation_length]),
438 correlation_length, correlation_scale);
439 int32_t energy2 = WebRtcSpl_DotProductWithScale(
440 &(audio_history[signal_length - correlation_length - best_index]),
441 &(audio_history[signal_length - correlation_length - best_index]),
442 correlation_length, correlation_scale);
444 // Calculate the correlation coefficient between the two portions of the
446 int16_t corr_coefficient;
447 if ((energy1 > 0) && (energy2 > 0)) {
448 int energy1_scale = std::max(16 - WebRtcSpl_NormW32(energy1), 0);
449 int energy2_scale = std::max(16 - WebRtcSpl_NormW32(energy2), 0);
450 // Make sure total scaling is even (to simplify scale factor after sqrt).
451 if ((energy1_scale + energy2_scale) & 1) {
452 // If sum is odd, add 1 to make it even.
455 int16_t scaled_energy1 = energy1 >> energy1_scale;
456 int16_t scaled_energy2 = energy2 >> energy2_scale;
457 int16_t sqrt_energy_product = WebRtcSpl_SqrtFloor(
458 scaled_energy1 * scaled_energy2);
459 // Calculate max_correlation / sqrt(energy1 * energy2) in Q14.
460 int cc_shift = 14 - (energy1_scale + energy2_scale) / 2;
461 max_correlation = WEBRTC_SPL_SHIFT_W32(max_correlation, cc_shift);
462 corr_coefficient = WebRtcSpl_DivW32W16(max_correlation,
463 sqrt_energy_product);
464 corr_coefficient = std::min(static_cast<int16_t>(16384),
465 corr_coefficient); // Cap at 1.0 in Q14.
467 corr_coefficient = 0;
470 // Extract the two vectors expand_vector0 and expand_vector1 from
472 int16_t expansion_length = static_cast<int16_t>(max_lag_ + overlap_length_);
473 const int16_t* vector1 = &(audio_history[signal_length - expansion_length]);
474 const int16_t* vector2 = vector1 - distortion_lag;
475 // Normalize the second vector to the same energy as the first.
476 energy1 = WebRtcSpl_DotProductWithScale(vector1, vector1, expansion_length,
478 energy2 = WebRtcSpl_DotProductWithScale(vector2, vector2, expansion_length,
480 // Confirm that amplitude ratio sqrt(energy1 / energy2) is within 0.5 - 2.0,
481 // i.e., energy1 / energy1 is within 0.25 - 4.
482 int16_t amplitude_ratio;
483 if ((energy1 / 4 < energy2) && (energy1 > energy2 / 4)) {
484 // Energy constraint fulfilled. Use both vectors and scale them
486 int16_t scaled_energy2 = std::max(16 - WebRtcSpl_NormW32(energy2), 0);
487 int16_t scaled_energy1 = scaled_energy2 - 13;
488 // Calculate scaled_energy1 / scaled_energy2 in Q13.
489 int32_t energy_ratio = WebRtcSpl_DivW32W16(
490 WEBRTC_SPL_SHIFT_W32(energy1, -scaled_energy1),
491 WEBRTC_SPL_RSHIFT_W32(energy2, scaled_energy2));
492 // Calculate sqrt ratio in Q13 (sqrt of en1/en2 in Q26).
493 amplitude_ratio = WebRtcSpl_SqrtFloor(energy_ratio << 13);
494 // Copy the two vectors and give them the same energy.
495 parameters.expand_vector0.Clear();
496 parameters.expand_vector0.PushBack(vector1, expansion_length);
497 parameters.expand_vector1.Clear();
498 if (parameters.expand_vector1.Size() <
499 static_cast<size_t>(expansion_length)) {
500 parameters.expand_vector1.Extend(
501 expansion_length - parameters.expand_vector1.Size());
503 WebRtcSpl_AffineTransformVector(¶meters.expand_vector1[0],
504 const_cast<int16_t*>(vector2),
510 // Energy change constraint not fulfilled. Only use last vector.
511 parameters.expand_vector0.Clear();
512 parameters.expand_vector0.PushBack(vector1, expansion_length);
513 // Copy from expand_vector0 to expand_vector1.
514 parameters.expand_vector0.CopyTo(¶meters.expand_vector1);
515 // Set the energy_ratio since it is used by muting slope.
516 if ((energy1 / 4 < energy2) || (energy2 == 0)) {
517 amplitude_ratio = 4096; // 0.5 in Q13.
519 amplitude_ratio = 16384; // 2.0 in Q13.
523 // Set the 3 lag values.
524 int lag_difference = distortion_lag - correlation_lag;
525 if (lag_difference == 0) {
526 // |distortion_lag| and |correlation_lag| are equal.
527 expand_lags_[0] = distortion_lag;
528 expand_lags_[1] = distortion_lag;
529 expand_lags_[2] = distortion_lag;
531 // |distortion_lag| and |correlation_lag| are not equal; use different
532 // combinations of the two.
533 // First lag is |distortion_lag| only.
534 expand_lags_[0] = distortion_lag;
535 // Second lag is the average of the two.
536 expand_lags_[1] = (distortion_lag + correlation_lag) / 2;
537 // Third lag is the average again, but rounding towards |correlation_lag|.
538 if (lag_difference > 0) {
539 expand_lags_[2] = (distortion_lag + correlation_lag - 1) / 2;
541 expand_lags_[2] = (distortion_lag + correlation_lag + 1) / 2;
545 // Calculate the LPC and the gain of the filters.
546 // Calculate scale value needed for auto-correlation.
547 correlation_scale = WebRtcSpl_MaxAbsValueW16(
548 &(audio_history[signal_length - fs_mult_lpc_analysis_len]),
549 fs_mult_lpc_analysis_len);
551 correlation_scale = std::min(16 - WebRtcSpl_NormW32(correlation_scale), 0);
552 correlation_scale = std::max(correlation_scale * 2 + 7, 0);
554 // Calculate kUnvoicedLpcOrder + 1 lags of the auto-correlation function.
555 size_t temp_index = signal_length - fs_mult_lpc_analysis_len -
557 // Copy signal to temporary vector to be able to pad with leading zeros.
558 int16_t* temp_signal = new int16_t[fs_mult_lpc_analysis_len
559 + kUnvoicedLpcOrder];
560 memset(temp_signal, 0,
561 sizeof(int16_t) * (fs_mult_lpc_analysis_len + kUnvoicedLpcOrder));
562 memcpy(&temp_signal[kUnvoicedLpcOrder],
563 &audio_history[temp_index + kUnvoicedLpcOrder],
564 sizeof(int16_t) * fs_mult_lpc_analysis_len);
565 WebRtcSpl_CrossCorrelation(auto_correlation,
566 &temp_signal[kUnvoicedLpcOrder],
567 &temp_signal[kUnvoicedLpcOrder],
568 fs_mult_lpc_analysis_len, kUnvoicedLpcOrder + 1,
569 correlation_scale, -1);
570 delete [] temp_signal;
572 // Verify that variance is positive.
573 if (auto_correlation[0] > 0) {
574 // Estimate AR filter parameters using Levinson-Durbin algorithm;
575 // kUnvoicedLpcOrder + 1 filter coefficients.
576 int16_t stability = WebRtcSpl_LevinsonDurbin(auto_correlation,
577 parameters.ar_filter,
581 // Keep filter parameters only if filter is stable.
582 if (stability != 1) {
583 // Set first coefficient to 4096 (1.0 in Q12).
584 parameters.ar_filter[0] = 4096;
585 // Set remaining |kUnvoicedLpcOrder| coefficients to zero.
586 WebRtcSpl_MemSetW16(parameters.ar_filter + 1, 0, kUnvoicedLpcOrder);
590 if (channel_ix == 0) {
591 // Extract a noise segment.
592 int16_t noise_length;
593 if (distortion_lag < 40) {
594 noise_length = 2 * distortion_lag + 30;
596 noise_length = distortion_lag + 30;
598 if (noise_length <= RandomVector::kRandomTableSize) {
599 memcpy(random_vector, RandomVector::kRandomTable,
600 sizeof(int16_t) * noise_length);
602 // Only applies to SWB where length could be larger than
603 // |kRandomTableSize|.
604 memcpy(random_vector, RandomVector::kRandomTable,
605 sizeof(int16_t) * RandomVector::kRandomTableSize);
606 assert(noise_length <= kMaxSampleRate / 8000 * 120 + 30);
607 random_vector_->IncreaseSeedIncrement(2);
608 random_vector_->Generate(
609 noise_length - RandomVector::kRandomTableSize,
610 &random_vector[RandomVector::kRandomTableSize]);
614 // Set up state vector and calculate scale factor for unvoiced filtering.
615 memcpy(parameters.ar_filter_state,
616 &(audio_history[signal_length - kUnvoicedLpcOrder]),
617 sizeof(int16_t) * kUnvoicedLpcOrder);
618 memcpy(unvoiced_vector - kUnvoicedLpcOrder,
619 &(audio_history[signal_length - 128 - kUnvoicedLpcOrder]),
620 sizeof(int16_t) * kUnvoicedLpcOrder);
621 WebRtcSpl_FilterMAFastQ12(
622 const_cast<int16_t*>(&audio_history[signal_length - 128]),
623 unvoiced_vector, parameters.ar_filter, kUnvoicedLpcOrder + 1, 128);
624 int16_t unvoiced_prescale;
625 if (WebRtcSpl_MaxAbsValueW16(unvoiced_vector, 128) > 4000) {
626 unvoiced_prescale = 4;
628 unvoiced_prescale = 0;
630 int32_t unvoiced_energy = WebRtcSpl_DotProductWithScale(unvoiced_vector,
635 // Normalize |unvoiced_energy| to 28 or 29 bits to preserve sqrt() accuracy.
636 int16_t unvoiced_scale = WebRtcSpl_NormW32(unvoiced_energy) - 3;
637 // Make sure we do an odd number of shifts since we already have 7 shifts
638 // from dividing with 128 earlier. This will make the total scale factor
639 // even, which is suitable for the sqrt.
640 unvoiced_scale += ((unvoiced_scale & 0x1) ^ 0x1);
641 unvoiced_energy = WEBRTC_SPL_SHIFT_W32(unvoiced_energy, unvoiced_scale);
642 int32_t unvoiced_gain = WebRtcSpl_SqrtFloor(unvoiced_energy);
643 parameters.ar_gain_scale = 13
644 + (unvoiced_scale + 7 - unvoiced_prescale) / 2;
645 parameters.ar_gain = unvoiced_gain;
647 // Calculate voice_mix_factor from corr_coefficient.
648 // Let x = corr_coefficient. Then, we compute:
650 // voice_mix_factor = (-5179 + 19931x - 16422x^2 + 5776x^3) / 4096;
652 // voice_mix_factor = 0;
653 if (corr_coefficient > 7875) {
655 x1 = corr_coefficient; // |corr_coefficient| is in Q14.
656 x2 = (x1 * x1) >> 14; // Shift 14 to keep result in Q14.
657 x3 = (x1 * x2) >> 14;
658 static const int kCoefficients[4] = { -5179, 19931, -16422, 5776 };
659 int32_t temp_sum = kCoefficients[0] << 14;
660 temp_sum += kCoefficients[1] * x1;
661 temp_sum += kCoefficients[2] * x2;
662 temp_sum += kCoefficients[3] * x3;
663 parameters.voice_mix_factor = temp_sum / 4096;
664 parameters.voice_mix_factor = std::min(parameters.voice_mix_factor,
665 static_cast<int16_t>(16384));
666 parameters.voice_mix_factor = std::max(parameters.voice_mix_factor,
667 static_cast<int16_t>(0));
669 parameters.voice_mix_factor = 0;
672 // Calculate muting slope. Reuse value from earlier scaling of
673 // |expand_vector0| and |expand_vector1|.
674 int16_t slope = amplitude_ratio;
677 // Calculate (1 - (1 / slope)) / distortion_lag =
678 // (slope - 1) / (distortion_lag * slope).
679 // |slope| is in Q13, so 1 corresponds to 8192. Shift up to Q25 before
681 // Shift the denominator from Q13 to Q5 before the division. The result of
682 // the division will then be in Q20.
683 int16_t temp_ratio = WebRtcSpl_DivW32W16((slope - 8192) << 12,
684 (distortion_lag * slope) >> 8);
687 // Divide by 2, with proper rounding.
688 parameters.mute_slope = (temp_ratio + 1) / 2;
690 // Divide by 8, with proper rounding.
691 parameters.mute_slope = (temp_ratio + 4) / 8;
693 parameters.onset = true;
695 // Calculate (1 - slope) / distortion_lag.
696 // Shift |slope| by 7 to Q20 before the division. The result is in Q20.
697 parameters.mute_slope = WebRtcSpl_DivW32W16((8192 - slope) << 7,
699 if (parameters.voice_mix_factor <= 13107) {
700 // Make sure the mute factor decreases from 1.0 to 0.9 in no more than
702 // mute_slope >= 0.005 / fs_mult in Q20.
703 parameters.mute_slope = std::max(static_cast<int16_t>(5243 / fs_mult),
704 parameters.mute_slope);
705 } else if (slope > 8028) {
706 parameters.mute_slope = 0;
708 parameters.onset = false;
713 int16_t Expand::Correlation(const int16_t* input, size_t input_length,
714 int16_t* output, int16_t* output_scale) const {
715 // Set parameters depending on sample rate.
716 const int16_t* filter_coefficients;
717 int16_t num_coefficients;
718 int16_t downsampling_factor;
719 if (fs_hz_ == 8000) {
720 num_coefficients = 3;
721 downsampling_factor = 2;
722 filter_coefficients = DspHelper::kDownsample8kHzTbl;
723 } else if (fs_hz_ == 16000) {
724 num_coefficients = 5;
725 downsampling_factor = 4;
726 filter_coefficients = DspHelper::kDownsample16kHzTbl;
727 } else if (fs_hz_ == 32000) {
728 num_coefficients = 7;
729 downsampling_factor = 8;
730 filter_coefficients = DspHelper::kDownsample32kHzTbl;
731 } else { // fs_hz_ == 48000.
732 num_coefficients = 7;
733 downsampling_factor = 12;
734 filter_coefficients = DspHelper::kDownsample48kHzTbl;
737 // Correlate from lag 10 to lag 60 in downsampled domain.
738 // (Corresponds to 20-120 for narrow-band, 40-240 for wide-band, and so on.)
739 static const int kCorrelationStartLag = 10;
740 static const int kNumCorrelationLags = 54;
741 static const int kCorrelationLength = 60;
742 // Downsample to 4 kHz sample rate.
743 static const int kDownsampledLength = kCorrelationStartLag
744 + kNumCorrelationLags + kCorrelationLength;
745 int16_t downsampled_input[kDownsampledLength];
746 static const int kFilterDelay = 0;
747 WebRtcSpl_DownsampleFast(
748 input + input_length - kDownsampledLength * downsampling_factor,
749 kDownsampledLength * downsampling_factor, downsampled_input,
750 kDownsampledLength, filter_coefficients, num_coefficients,
751 downsampling_factor, kFilterDelay);
753 // Normalize |downsampled_input| to using all 16 bits.
754 int16_t max_value = WebRtcSpl_MaxAbsValueW16(downsampled_input,
756 int16_t norm_shift = 16 - WebRtcSpl_NormW32(max_value);
757 WebRtcSpl_VectorBitShiftW16(downsampled_input, kDownsampledLength,
758 downsampled_input, norm_shift);
760 int32_t correlation[kNumCorrelationLags];
761 static const int kCorrelationShift = 6;
762 WebRtcSpl_CrossCorrelation(
764 &downsampled_input[kDownsampledLength - kCorrelationLength],
765 &downsampled_input[kDownsampledLength - kCorrelationLength
766 - kCorrelationStartLag],
767 kCorrelationLength, kNumCorrelationLags, kCorrelationShift, -1);
769 // Normalize and move data from 32-bit to 16-bit vector.
770 int32_t max_correlation = WebRtcSpl_MaxAbsValueW32(correlation,
771 kNumCorrelationLags);
772 int16_t norm_shift2 = std::max(18 - WebRtcSpl_NormW32(max_correlation), 0);
773 WebRtcSpl_VectorBitShiftW32ToW16(output, kNumCorrelationLags, correlation,
775 // Total scale factor (right shifts) of correlation value.
776 *output_scale = 2 * norm_shift + kCorrelationShift + norm_shift2;
777 return kNumCorrelationLags;
780 void Expand::UpdateLagIndex() {
781 current_lag_index_ = current_lag_index_ + lag_index_direction_;
782 // Change direction if needed.
783 if (current_lag_index_ <= 0) {
784 lag_index_direction_ = 1;
786 if (current_lag_index_ >= kNumLags - 1) {
787 lag_index_direction_ = -1;
791 Expand* ExpandFactory::Create(BackgroundNoise* background_noise,
792 SyncBuffer* sync_buffer,
793 RandomVector* random_vector,
795 size_t num_channels) const {
796 return new Expand(background_noise, sync_buffer, random_vector, fs,
800 // TODO(turajs): This can be moved to BackgroundNoise class.
801 void Expand::GenerateBackgroundNoise(int16_t* random_vector,
804 bool too_many_expands,
805 size_t num_noise_samples,
807 static const int kNoiseLpcOrder = BackgroundNoise::kMaxLpcOrder;
808 int16_t scaled_random_vector[kMaxSampleRate / 8000 * 125];
809 assert(static_cast<size_t>(kMaxSampleRate / 8000 * 125) >= num_noise_samples);
810 int16_t* noise_samples = &buffer[kNoiseLpcOrder];
811 if (background_noise_->initialized()) {
812 // Use background noise parameters.
813 memcpy(noise_samples - kNoiseLpcOrder,
814 background_noise_->FilterState(channel),
815 sizeof(int16_t) * kNoiseLpcOrder);
818 if (background_noise_->ScaleShift(channel) > 1) {
819 dc_offset = 1 << (background_noise_->ScaleShift(channel) - 1);
822 // Scale random vector to correct energy level.
823 WebRtcSpl_AffineTransformVector(
824 scaled_random_vector, random_vector,
825 background_noise_->Scale(channel), dc_offset,
826 background_noise_->ScaleShift(channel),
827 static_cast<int>(num_noise_samples));
829 WebRtcSpl_FilterARFastQ12(scaled_random_vector, noise_samples,
830 background_noise_->Filter(channel),
832 static_cast<int>(num_noise_samples));
834 background_noise_->SetFilterState(
836 &(noise_samples[num_noise_samples - kNoiseLpcOrder]),
839 // Unmute the background noise.
840 int16_t bgn_mute_factor = background_noise_->MuteFactor(channel);
841 NetEq::BackgroundNoiseMode bgn_mode = background_noise_->mode();
842 if (bgn_mode == NetEq::kBgnFade && too_many_expands &&
843 bgn_mute_factor > 0) {
845 // Calculate muting slope, approximately -2^18 / fs_hz.
847 if (fs_hz_ == 8000) {
849 } else if (fs_hz_ == 16000) {
851 } else if (fs_hz_ == 32000) {
856 // Use UnmuteSignal function with negative slope.
857 // |bgn_mute_factor| is in Q14. |mute_slope| is in Q20.
858 DspHelper::UnmuteSignal(noise_samples,
863 } else if (bgn_mute_factor < 16384) {
864 // If mode is kBgnOn, or if kBgnFade has started fading,
865 // use regular |mute_slope|.
866 if (!stop_muting_ && bgn_mode != NetEq::kBgnOff &&
867 !(bgn_mode == NetEq::kBgnFade && too_many_expands)) {
868 DspHelper::UnmuteSignal(noise_samples,
869 static_cast<int>(num_noise_samples),
874 // kBgnOn and stop muting, or
875 // kBgnOff (mute factor is always 0), or
876 // kBgnFade has reached 0.
877 WebRtcSpl_AffineTransformVector(noise_samples, noise_samples,
878 bgn_mute_factor, 8192, 14,
879 static_cast<int>(num_noise_samples));
882 // Update mute_factor in BackgroundNoise class.
883 background_noise_->SetMuteFactor(channel, bgn_mute_factor);
885 // BGN parameters have not been initialized; use zero noise.
886 memset(noise_samples, 0, sizeof(int16_t) * num_noise_samples);
890 void Expand::GenerateRandomVector(int seed_increment,
892 int16_t* random_vector) {
893 // TODO(turajs): According to hlundin The loop should not be needed. Should be
894 // just as good to generate all of the vector in one call.
895 size_t samples_generated = 0;
896 const size_t kMaxRandSamples = RandomVector::kRandomTableSize;
897 while (samples_generated < length) {
898 size_t rand_length = std::min(length - samples_generated, kMaxRandSamples);
899 random_vector_->IncreaseSeedIncrement(seed_increment);
900 random_vector_->Generate(rand_length, &random_vector[samples_generated]);
901 samples_generated += rand_length;
905 } // namespace webrtc