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 "modules/audio_coding/neteq/expand.h"
12
13 #include <assert.h>
14 #include <string.h> // memset
15
16 #include <algorithm> // min, max
17 #include <limits> // numeric_limits<T>
18
19 #include "common_audio/signal_processing/include/signal_processing_library.h"
20 #include "modules/audio_coding/neteq/background_noise.h"
21 #include "modules/audio_coding/neteq/cross_correlation.h"
22 #include "modules/audio_coding/neteq/dsp_helper.h"
23 #include "modules/audio_coding/neteq/random_vector.h"
24 #include "modules/audio_coding/neteq/statistics_calculator.h"
25 #include "modules/audio_coding/neteq/sync_buffer.h"
26 #include "rtc_base/numerics/safe_conversions.h"
27
28 namespace webrtc {
29
Expand(BackgroundNoise * background_noise,SyncBuffer * sync_buffer,RandomVector * random_vector,StatisticsCalculator * statistics,int fs,size_t num_channels)30 Expand::Expand(BackgroundNoise* background_noise,
31 SyncBuffer* sync_buffer,
32 RandomVector* random_vector,
33 StatisticsCalculator* statistics,
34 int fs,
35 size_t num_channels)
36 : random_vector_(random_vector),
37 sync_buffer_(sync_buffer),
38 first_expand_(true),
39 fs_hz_(fs),
40 num_channels_(num_channels),
41 consecutive_expands_(0),
42 background_noise_(background_noise),
43 statistics_(statistics),
44 overlap_length_(5 * fs / 8000),
45 lag_index_direction_(0),
46 current_lag_index_(0),
47 stop_muting_(false),
48 expand_duration_samples_(0),
49 channel_parameters_(new ChannelParameters[num_channels_]) {
50 assert(fs == 8000 || fs == 16000 || fs == 32000 || fs == 48000);
51 assert(fs <= static_cast<int>(kMaxSampleRate)); // Should not be possible.
52 assert(num_channels_ > 0);
53 memset(expand_lags_, 0, sizeof(expand_lags_));
54 Reset();
55 }
56
57 Expand::~Expand() = default;
58
Reset()59 void Expand::Reset() {
60 first_expand_ = true;
61 consecutive_expands_ = 0;
62 max_lag_ = 0;
63 for (size_t ix = 0; ix < num_channels_; ++ix) {
64 channel_parameters_[ix].expand_vector0.Clear();
65 channel_parameters_[ix].expand_vector1.Clear();
66 }
67 }
68
Process(AudioMultiVector * output)69 int Expand::Process(AudioMultiVector* output) {
70 int16_t random_vector[kMaxSampleRate / 8000 * 120 + 30];
71 int16_t scaled_random_vector[kMaxSampleRate / 8000 * 125];
72 static const int kTempDataSize = 3600;
73 int16_t temp_data[kTempDataSize]; // TODO(hlundin) Remove this.
74 int16_t* voiced_vector_storage = temp_data;
75 int16_t* voiced_vector = &voiced_vector_storage[overlap_length_];
76 static const size_t kNoiseLpcOrder = BackgroundNoise::kMaxLpcOrder;
77 int16_t unvoiced_array_memory[kNoiseLpcOrder + kMaxSampleRate / 8000 * 125];
78 int16_t* unvoiced_vector = unvoiced_array_memory + kUnvoicedLpcOrder;
79 int16_t* noise_vector = unvoiced_array_memory + kNoiseLpcOrder;
80
81 int fs_mult = fs_hz_ / 8000;
82
83 if (first_expand_) {
84 // Perform initial setup if this is the first expansion since last reset.
85 AnalyzeSignal(random_vector);
86 first_expand_ = false;
87 expand_duration_samples_ = 0;
88 } else {
89 // This is not the first expansion, parameters are already estimated.
90 // Extract a noise segment.
91 size_t rand_length = max_lag_;
92 // This only applies to SWB where length could be larger than 256.
93 assert(rand_length <= kMaxSampleRate / 8000 * 120 + 30);
94 GenerateRandomVector(2, rand_length, random_vector);
95 }
96
97
98 // Generate signal.
99 UpdateLagIndex();
100
101 // Voiced part.
102 // Generate a weighted vector with the current lag.
103 size_t expansion_vector_length = max_lag_ + overlap_length_;
104 size_t current_lag = expand_lags_[current_lag_index_];
105 // Copy lag+overlap data.
106 size_t expansion_vector_position = expansion_vector_length - current_lag -
107 overlap_length_;
108 size_t temp_length = current_lag + overlap_length_;
109 for (size_t channel_ix = 0; channel_ix < num_channels_; ++channel_ix) {
110 ChannelParameters& parameters = channel_parameters_[channel_ix];
111 if (current_lag_index_ == 0) {
112 // Use only expand_vector0.
113 assert(expansion_vector_position + temp_length <=
114 parameters.expand_vector0.Size());
115 parameters.expand_vector0.CopyTo(temp_length, expansion_vector_position,
116 voiced_vector_storage);
117 } else if (current_lag_index_ == 1) {
118 std::unique_ptr<int16_t[]> temp_0(new int16_t[temp_length]);
119 parameters.expand_vector0.CopyTo(temp_length, expansion_vector_position,
120 temp_0.get());
121 std::unique_ptr<int16_t[]> temp_1(new int16_t[temp_length]);
122 parameters.expand_vector1.CopyTo(temp_length, expansion_vector_position,
123 temp_1.get());
124 // Mix 3/4 of expand_vector0 with 1/4 of expand_vector1.
125 WebRtcSpl_ScaleAndAddVectorsWithRound(temp_0.get(), 3, temp_1.get(), 1, 2,
126 voiced_vector_storage, temp_length);
127 } else if (current_lag_index_ == 2) {
128 // Mix 1/2 of expand_vector0 with 1/2 of expand_vector1.
129 assert(expansion_vector_position + temp_length <=
130 parameters.expand_vector0.Size());
131 assert(expansion_vector_position + temp_length <=
132 parameters.expand_vector1.Size());
133
134 std::unique_ptr<int16_t[]> temp_0(new int16_t[temp_length]);
135 parameters.expand_vector0.CopyTo(temp_length, expansion_vector_position,
136 temp_0.get());
137 std::unique_ptr<int16_t[]> temp_1(new int16_t[temp_length]);
138 parameters.expand_vector1.CopyTo(temp_length, expansion_vector_position,
139 temp_1.get());
140 WebRtcSpl_ScaleAndAddVectorsWithRound(temp_0.get(), 1, temp_1.get(), 1, 1,
141 voiced_vector_storage, temp_length);
142 }
143
144 // Get tapering window parameters. Values are in Q15.
145 int16_t muting_window, muting_window_increment;
146 int16_t unmuting_window, unmuting_window_increment;
147 if (fs_hz_ == 8000) {
148 muting_window = DspHelper::kMuteFactorStart8kHz;
149 muting_window_increment = DspHelper::kMuteFactorIncrement8kHz;
150 unmuting_window = DspHelper::kUnmuteFactorStart8kHz;
151 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement8kHz;
152 } else if (fs_hz_ == 16000) {
153 muting_window = DspHelper::kMuteFactorStart16kHz;
154 muting_window_increment = DspHelper::kMuteFactorIncrement16kHz;
155 unmuting_window = DspHelper::kUnmuteFactorStart16kHz;
156 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement16kHz;
157 } else if (fs_hz_ == 32000) {
158 muting_window = DspHelper::kMuteFactorStart32kHz;
159 muting_window_increment = DspHelper::kMuteFactorIncrement32kHz;
160 unmuting_window = DspHelper::kUnmuteFactorStart32kHz;
161 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement32kHz;
162 } else { // fs_ == 48000
163 muting_window = DspHelper::kMuteFactorStart48kHz;
164 muting_window_increment = DspHelper::kMuteFactorIncrement48kHz;
165 unmuting_window = DspHelper::kUnmuteFactorStart48kHz;
166 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement48kHz;
167 }
168
169 // Smooth the expanded if it has not been muted to a low amplitude and
170 // |current_voice_mix_factor| is larger than 0.5.
171 if ((parameters.mute_factor > 819) &&
172 (parameters.current_voice_mix_factor > 8192)) {
173 size_t start_ix = sync_buffer_->Size() - overlap_length_;
174 for (size_t i = 0; i < overlap_length_; i++) {
175 // Do overlap add between new vector and overlap.
176 (*sync_buffer_)[channel_ix][start_ix + i] =
177 (((*sync_buffer_)[channel_ix][start_ix + i] * muting_window) +
178 (((parameters.mute_factor * voiced_vector_storage[i]) >> 14) *
179 unmuting_window) + 16384) >> 15;
180 muting_window += muting_window_increment;
181 unmuting_window += unmuting_window_increment;
182 }
183 } else if (parameters.mute_factor == 0) {
184 // The expanded signal will consist of only comfort noise if
185 // mute_factor = 0. Set the output length to 15 ms for best noise
186 // production.
187 // TODO(hlundin): This has been disabled since the length of
188 // parameters.expand_vector0 and parameters.expand_vector1 no longer
189 // match with expand_lags_, causing invalid reads and writes. Is it a good
190 // idea to enable this again, and solve the vector size problem?
191 // max_lag_ = fs_mult * 120;
192 // expand_lags_[0] = fs_mult * 120;
193 // expand_lags_[1] = fs_mult * 120;
194 // expand_lags_[2] = fs_mult * 120;
195 }
196
197 // Unvoiced part.
198 // Filter |scaled_random_vector| through |ar_filter_|.
199 memcpy(unvoiced_vector - kUnvoicedLpcOrder, parameters.ar_filter_state,
200 sizeof(int16_t) * kUnvoicedLpcOrder);
201 int32_t add_constant = 0;
202 if (parameters.ar_gain_scale > 0) {
203 add_constant = 1 << (parameters.ar_gain_scale - 1);
204 }
205 WebRtcSpl_AffineTransformVector(scaled_random_vector, random_vector,
206 parameters.ar_gain, add_constant,
207 parameters.ar_gain_scale,
208 current_lag);
209 WebRtcSpl_FilterARFastQ12(scaled_random_vector, unvoiced_vector,
210 parameters.ar_filter, kUnvoicedLpcOrder + 1,
211 current_lag);
212 memcpy(parameters.ar_filter_state,
213 &(unvoiced_vector[current_lag - kUnvoicedLpcOrder]),
214 sizeof(int16_t) * kUnvoicedLpcOrder);
215
216 // Combine voiced and unvoiced contributions.
217
218 // Set a suitable cross-fading slope.
219 // For lag =
220 // <= 31 * fs_mult => go from 1 to 0 in about 8 ms;
221 // (>= 31 .. <= 63) * fs_mult => go from 1 to 0 in about 16 ms;
222 // >= 64 * fs_mult => go from 1 to 0 in about 32 ms.
223 // temp_shift = getbits(max_lag_) - 5.
224 int temp_shift =
225 (31 - WebRtcSpl_NormW32(rtc::dchecked_cast<int32_t>(max_lag_))) - 5;
226 int16_t mix_factor_increment = 256 >> temp_shift;
227 if (stop_muting_) {
228 mix_factor_increment = 0;
229 }
230
231 // Create combined signal by shifting in more and more of unvoiced part.
232 temp_shift = 8 - temp_shift; // = getbits(mix_factor_increment).
233 size_t temp_length = (parameters.current_voice_mix_factor -
234 parameters.voice_mix_factor) >> temp_shift;
235 temp_length = std::min(temp_length, current_lag);
236 DspHelper::CrossFade(voiced_vector, unvoiced_vector, temp_length,
237 ¶meters.current_voice_mix_factor,
238 mix_factor_increment, temp_data);
239
240 // End of cross-fading period was reached before end of expanded signal
241 // path. Mix the rest with a fixed mixing factor.
242 if (temp_length < current_lag) {
243 if (mix_factor_increment != 0) {
244 parameters.current_voice_mix_factor = parameters.voice_mix_factor;
245 }
246 int16_t temp_scale = 16384 - parameters.current_voice_mix_factor;
247 WebRtcSpl_ScaleAndAddVectorsWithRound(
248 voiced_vector + temp_length, parameters.current_voice_mix_factor,
249 unvoiced_vector + temp_length, temp_scale, 14,
250 temp_data + temp_length, current_lag - temp_length);
251 }
252
253 // Select muting slope depending on how many consecutive expands we have
254 // done.
255 if (consecutive_expands_ == 3) {
256 // Let the mute factor decrease from 1.0 to 0.95 in 6.25 ms.
257 // mute_slope = 0.0010 / fs_mult in Q20.
258 parameters.mute_slope = std::max(parameters.mute_slope, 1049 / fs_mult);
259 }
260 if (consecutive_expands_ == 7) {
261 // Let the mute factor decrease from 1.0 to 0.90 in 6.25 ms.
262 // mute_slope = 0.0020 / fs_mult in Q20.
263 parameters.mute_slope = std::max(parameters.mute_slope, 2097 / fs_mult);
264 }
265
266 // Mute segment according to slope value.
267 if ((consecutive_expands_ != 0) || !parameters.onset) {
268 // Mute to the previous level, then continue with the muting.
269 WebRtcSpl_AffineTransformVector(temp_data, temp_data,
270 parameters.mute_factor, 8192,
271 14, current_lag);
272
273 if (!stop_muting_) {
274 DspHelper::MuteSignal(temp_data, parameters.mute_slope, current_lag);
275
276 // Shift by 6 to go from Q20 to Q14.
277 // TODO(hlundin): Adding 8192 before shifting 6 steps seems wrong.
278 // Legacy.
279 int16_t gain = static_cast<int16_t>(16384 -
280 (((current_lag * parameters.mute_slope) + 8192) >> 6));
281 gain = ((gain * parameters.mute_factor) + 8192) >> 14;
282
283 // Guard against getting stuck with very small (but sometimes audible)
284 // gain.
285 if ((consecutive_expands_ > 3) && (gain >= parameters.mute_factor)) {
286 parameters.mute_factor = 0;
287 } else {
288 parameters.mute_factor = gain;
289 }
290 }
291 }
292
293 // Background noise part.
294 GenerateBackgroundNoise(random_vector,
295 channel_ix,
296 channel_parameters_[channel_ix].mute_slope,
297 TooManyExpands(),
298 current_lag,
299 unvoiced_array_memory);
300
301 // Add background noise to the combined voiced-unvoiced signal.
302 for (size_t i = 0; i < current_lag; i++) {
303 temp_data[i] = temp_data[i] + noise_vector[i];
304 }
305 if (channel_ix == 0) {
306 output->AssertSize(current_lag);
307 } else {
308 assert(output->Size() == current_lag);
309 }
310 (*output)[channel_ix].OverwriteAt(temp_data, current_lag, 0);
311 }
312
313 // Increase call number and cap it.
314 consecutive_expands_ = consecutive_expands_ >= kMaxConsecutiveExpands ?
315 kMaxConsecutiveExpands : consecutive_expands_ + 1;
316 expand_duration_samples_ += output->Size();
317 // Clamp the duration counter at 2 seconds.
318 expand_duration_samples_ = std::min(expand_duration_samples_,
319 rtc::dchecked_cast<size_t>(fs_hz_ * 2));
320 return 0;
321 }
322
SetParametersForNormalAfterExpand()323 void Expand::SetParametersForNormalAfterExpand() {
324 current_lag_index_ = 0;
325 lag_index_direction_ = 0;
326 stop_muting_ = true; // Do not mute signal any more.
327 statistics_->LogDelayedPacketOutageEvent(
328 rtc::dchecked_cast<int>(expand_duration_samples_) / (fs_hz_ / 1000));
329 }
330
SetParametersForMergeAfterExpand()331 void Expand::SetParametersForMergeAfterExpand() {
332 current_lag_index_ = -1; /* out of the 3 possible ones */
333 lag_index_direction_ = 1; /* make sure we get the "optimal" lag */
334 stop_muting_ = true;
335 }
336
Muted() const337 bool Expand::Muted() const {
338 if (first_expand_ || stop_muting_)
339 return false;
340 RTC_DCHECK(channel_parameters_);
341 for (size_t ch = 0; ch < num_channels_; ++ch) {
342 if (channel_parameters_[ch].mute_factor != 0)
343 return false;
344 }
345 return true;
346 }
347
overlap_length() const348 size_t Expand::overlap_length() const {
349 return overlap_length_;
350 }
351
InitializeForAnExpandPeriod()352 void Expand::InitializeForAnExpandPeriod() {
353 lag_index_direction_ = 1;
354 current_lag_index_ = -1;
355 stop_muting_ = false;
356 random_vector_->set_seed_increment(1);
357 consecutive_expands_ = 0;
358 for (size_t ix = 0; ix < num_channels_; ++ix) {
359 channel_parameters_[ix].current_voice_mix_factor = 16384; // 1.0 in Q14.
360 channel_parameters_[ix].mute_factor = 16384; // 1.0 in Q14.
361 // Start with 0 gain for background noise.
362 background_noise_->SetMuteFactor(ix, 0);
363 }
364 }
365
TooManyExpands()366 bool Expand::TooManyExpands() {
367 return consecutive_expands_ >= kMaxConsecutiveExpands;
368 }
369
AnalyzeSignal(int16_t * random_vector)370 void Expand::AnalyzeSignal(int16_t* random_vector) {
371 int32_t auto_correlation[kUnvoicedLpcOrder + 1];
372 int16_t reflection_coeff[kUnvoicedLpcOrder];
373 int16_t correlation_vector[kMaxSampleRate / 8000 * 102];
374 size_t best_correlation_index[kNumCorrelationCandidates];
375 int16_t best_correlation[kNumCorrelationCandidates];
376 size_t best_distortion_index[kNumCorrelationCandidates];
377 int16_t best_distortion[kNumCorrelationCandidates];
378 int32_t correlation_vector2[(99 * kMaxSampleRate / 8000) + 1];
379 int32_t best_distortion_w32[kNumCorrelationCandidates];
380 static const size_t kNoiseLpcOrder = BackgroundNoise::kMaxLpcOrder;
381 int16_t unvoiced_array_memory[kNoiseLpcOrder + kMaxSampleRate / 8000 * 125];
382 int16_t* unvoiced_vector = unvoiced_array_memory + kUnvoicedLpcOrder;
383
384 int fs_mult = fs_hz_ / 8000;
385
386 // Pre-calculate common multiplications with fs_mult.
387 size_t fs_mult_4 = static_cast<size_t>(fs_mult * 4);
388 size_t fs_mult_20 = static_cast<size_t>(fs_mult * 20);
389 size_t fs_mult_120 = static_cast<size_t>(fs_mult * 120);
390 size_t fs_mult_dist_len = fs_mult * kDistortionLength;
391 size_t fs_mult_lpc_analysis_len = fs_mult * kLpcAnalysisLength;
392
393 const size_t signal_length = static_cast<size_t>(256 * fs_mult);
394
395 const size_t audio_history_position = sync_buffer_->Size() - signal_length;
396 std::unique_ptr<int16_t[]> audio_history(new int16_t[signal_length]);
397 (*sync_buffer_)[0].CopyTo(signal_length, audio_history_position,
398 audio_history.get());
399
400 // Initialize.
401 InitializeForAnExpandPeriod();
402
403 // Calculate correlation in downsampled domain (4 kHz sample rate).
404 size_t correlation_length = 51; // TODO(hlundin): Legacy bit-exactness.
405 // If it is decided to break bit-exactness |correlation_length| should be
406 // initialized to the return value of Correlation().
407 Correlation(audio_history.get(), signal_length, correlation_vector);
408
409 // Find peaks in correlation vector.
410 DspHelper::PeakDetection(correlation_vector, correlation_length,
411 kNumCorrelationCandidates, fs_mult,
412 best_correlation_index, best_correlation);
413
414 // Adjust peak locations; cross-correlation lags start at 2.5 ms
415 // (20 * fs_mult samples).
416 best_correlation_index[0] += fs_mult_20;
417 best_correlation_index[1] += fs_mult_20;
418 best_correlation_index[2] += fs_mult_20;
419
420 // Calculate distortion around the |kNumCorrelationCandidates| best lags.
421 int distortion_scale = 0;
422 for (size_t i = 0; i < kNumCorrelationCandidates; i++) {
423 size_t min_index = std::max(fs_mult_20,
424 best_correlation_index[i] - fs_mult_4);
425 size_t max_index = std::min(fs_mult_120 - 1,
426 best_correlation_index[i] + fs_mult_4);
427 best_distortion_index[i] = DspHelper::MinDistortion(
428 &(audio_history[signal_length - fs_mult_dist_len]), min_index,
429 max_index, fs_mult_dist_len, &best_distortion_w32[i]);
430 distortion_scale = std::max(16 - WebRtcSpl_NormW32(best_distortion_w32[i]),
431 distortion_scale);
432 }
433 // Shift the distortion values to fit in 16 bits.
434 WebRtcSpl_VectorBitShiftW32ToW16(best_distortion, kNumCorrelationCandidates,
435 best_distortion_w32, distortion_scale);
436
437 // Find the maximizing index |i| of the cost function
438 // f[i] = best_correlation[i] / best_distortion[i].
439 int32_t best_ratio = std::numeric_limits<int32_t>::min();
440 size_t best_index = std::numeric_limits<size_t>::max();
441 for (size_t i = 0; i < kNumCorrelationCandidates; ++i) {
442 int32_t ratio;
443 if (best_distortion[i] > 0) {
444 ratio = (best_correlation[i] * (1 << 16)) / best_distortion[i];
445 } else if (best_correlation[i] == 0) {
446 ratio = 0; // No correlation set result to zero.
447 } else {
448 ratio = std::numeric_limits<int32_t>::max(); // Denominator is zero.
449 }
450 if (ratio > best_ratio) {
451 best_index = i;
452 best_ratio = ratio;
453 }
454 }
455
456 size_t distortion_lag = best_distortion_index[best_index];
457 size_t correlation_lag = best_correlation_index[best_index];
458 max_lag_ = std::max(distortion_lag, correlation_lag);
459
460 // Calculate the exact best correlation in the range between
461 // |correlation_lag| and |distortion_lag|.
462 correlation_length =
463 std::max(std::min(distortion_lag + 10, fs_mult_120),
464 static_cast<size_t>(60 * fs_mult));
465
466 size_t start_index = std::min(distortion_lag, correlation_lag);
467 size_t correlation_lags = static_cast<size_t>(
468 WEBRTC_SPL_ABS_W16((distortion_lag-correlation_lag)) + 1);
469 assert(correlation_lags <= static_cast<size_t>(99 * fs_mult + 1));
470
471 for (size_t channel_ix = 0; channel_ix < num_channels_; ++channel_ix) {
472 ChannelParameters& parameters = channel_parameters_[channel_ix];
473 // Calculate suitable scaling.
474 int16_t signal_max = WebRtcSpl_MaxAbsValueW16(
475 &audio_history[signal_length - correlation_length - start_index
476 - correlation_lags],
477 correlation_length + start_index + correlation_lags - 1);
478 int correlation_scale = (31 - WebRtcSpl_NormW32(signal_max * signal_max)) +
479 (31 - WebRtcSpl_NormW32(static_cast<int32_t>(correlation_length))) - 31;
480 correlation_scale = std::max(0, correlation_scale);
481
482 // Calculate the correlation, store in |correlation_vector2|.
483 WebRtcSpl_CrossCorrelation(
484 correlation_vector2,
485 &(audio_history[signal_length - correlation_length]),
486 &(audio_history[signal_length - correlation_length - start_index]),
487 correlation_length, correlation_lags, correlation_scale, -1);
488
489 // Find maximizing index.
490 best_index = WebRtcSpl_MaxIndexW32(correlation_vector2, correlation_lags);
491 int32_t max_correlation = correlation_vector2[best_index];
492 // Compensate index with start offset.
493 best_index = best_index + start_index;
494
495 // Calculate energies.
496 int32_t energy1 = WebRtcSpl_DotProductWithScale(
497 &(audio_history[signal_length - correlation_length]),
498 &(audio_history[signal_length - correlation_length]),
499 correlation_length, correlation_scale);
500 int32_t energy2 = WebRtcSpl_DotProductWithScale(
501 &(audio_history[signal_length - correlation_length - best_index]),
502 &(audio_history[signal_length - correlation_length - best_index]),
503 correlation_length, correlation_scale);
504
505 // Calculate the correlation coefficient between the two portions of the
506 // signal.
507 int32_t corr_coefficient;
508 if ((energy1 > 0) && (energy2 > 0)) {
509 int energy1_scale = std::max(16 - WebRtcSpl_NormW32(energy1), 0);
510 int energy2_scale = std::max(16 - WebRtcSpl_NormW32(energy2), 0);
511 // Make sure total scaling is even (to simplify scale factor after sqrt).
512 if ((energy1_scale + energy2_scale) & 1) {
513 // If sum is odd, add 1 to make it even.
514 energy1_scale += 1;
515 }
516 int32_t scaled_energy1 = energy1 >> energy1_scale;
517 int32_t scaled_energy2 = energy2 >> energy2_scale;
518 int16_t sqrt_energy_product = static_cast<int16_t>(
519 WebRtcSpl_SqrtFloor(scaled_energy1 * scaled_energy2));
520 // Calculate max_correlation / sqrt(energy1 * energy2) in Q14.
521 int cc_shift = 14 - (energy1_scale + energy2_scale) / 2;
522 max_correlation = WEBRTC_SPL_SHIFT_W32(max_correlation, cc_shift);
523 corr_coefficient = WebRtcSpl_DivW32W16(max_correlation,
524 sqrt_energy_product);
525 // Cap at 1.0 in Q14.
526 corr_coefficient = std::min(16384, corr_coefficient);
527 } else {
528 corr_coefficient = 0;
529 }
530
531 // Extract the two vectors expand_vector0 and expand_vector1 from
532 // |audio_history|.
533 size_t expansion_length = max_lag_ + overlap_length_;
534 const int16_t* vector1 = &(audio_history[signal_length - expansion_length]);
535 const int16_t* vector2 = vector1 - distortion_lag;
536 // Normalize the second vector to the same energy as the first.
537 energy1 = WebRtcSpl_DotProductWithScale(vector1, vector1, expansion_length,
538 correlation_scale);
539 energy2 = WebRtcSpl_DotProductWithScale(vector2, vector2, expansion_length,
540 correlation_scale);
541 // Confirm that amplitude ratio sqrt(energy1 / energy2) is within 0.5 - 2.0,
542 // i.e., energy1 / energy2 is within 0.25 - 4.
543 int16_t amplitude_ratio;
544 if ((energy1 / 4 < energy2) && (energy1 > energy2 / 4)) {
545 // Energy constraint fulfilled. Use both vectors and scale them
546 // accordingly.
547 int32_t scaled_energy2 = std::max(16 - WebRtcSpl_NormW32(energy2), 0);
548 int32_t scaled_energy1 = scaled_energy2 - 13;
549 // Calculate scaled_energy1 / scaled_energy2 in Q13.
550 int32_t energy_ratio = WebRtcSpl_DivW32W16(
551 WEBRTC_SPL_SHIFT_W32(energy1, -scaled_energy1),
552 static_cast<int16_t>(energy2 >> scaled_energy2));
553 // Calculate sqrt ratio in Q13 (sqrt of en1/en2 in Q26).
554 amplitude_ratio =
555 static_cast<int16_t>(WebRtcSpl_SqrtFloor(energy_ratio << 13));
556 // Copy the two vectors and give them the same energy.
557 parameters.expand_vector0.Clear();
558 parameters.expand_vector0.PushBack(vector1, expansion_length);
559 parameters.expand_vector1.Clear();
560 if (parameters.expand_vector1.Size() < expansion_length) {
561 parameters.expand_vector1.Extend(
562 expansion_length - parameters.expand_vector1.Size());
563 }
564 std::unique_ptr<int16_t[]> temp_1(new int16_t[expansion_length]);
565 WebRtcSpl_AffineTransformVector(temp_1.get(),
566 const_cast<int16_t*>(vector2),
567 amplitude_ratio,
568 4096,
569 13,
570 expansion_length);
571 parameters.expand_vector1.OverwriteAt(temp_1.get(), expansion_length, 0);
572 } else {
573 // Energy change constraint not fulfilled. Only use last vector.
574 parameters.expand_vector0.Clear();
575 parameters.expand_vector0.PushBack(vector1, expansion_length);
576 // Copy from expand_vector0 to expand_vector1.
577 parameters.expand_vector0.CopyTo(¶meters.expand_vector1);
578 // Set the energy_ratio since it is used by muting slope.
579 if ((energy1 / 4 < energy2) || (energy2 == 0)) {
580 amplitude_ratio = 4096; // 0.5 in Q13.
581 } else {
582 amplitude_ratio = 16384; // 2.0 in Q13.
583 }
584 }
585
586 // Set the 3 lag values.
587 if (distortion_lag == correlation_lag) {
588 expand_lags_[0] = distortion_lag;
589 expand_lags_[1] = distortion_lag;
590 expand_lags_[2] = distortion_lag;
591 } else {
592 // |distortion_lag| and |correlation_lag| are not equal; use different
593 // combinations of the two.
594 // First lag is |distortion_lag| only.
595 expand_lags_[0] = distortion_lag;
596 // Second lag is the average of the two.
597 expand_lags_[1] = (distortion_lag + correlation_lag) / 2;
598 // Third lag is the average again, but rounding towards |correlation_lag|.
599 if (distortion_lag > correlation_lag) {
600 expand_lags_[2] = (distortion_lag + correlation_lag - 1) / 2;
601 } else {
602 expand_lags_[2] = (distortion_lag + correlation_lag + 1) / 2;
603 }
604 }
605
606 // Calculate the LPC and the gain of the filters.
607
608 // Calculate kUnvoicedLpcOrder + 1 lags of the auto-correlation function.
609 size_t temp_index = signal_length - fs_mult_lpc_analysis_len -
610 kUnvoicedLpcOrder;
611 // Copy signal to temporary vector to be able to pad with leading zeros.
612 int16_t* temp_signal = new int16_t[fs_mult_lpc_analysis_len
613 + kUnvoicedLpcOrder];
614 memset(temp_signal, 0,
615 sizeof(int16_t) * (fs_mult_lpc_analysis_len + kUnvoicedLpcOrder));
616 memcpy(&temp_signal[kUnvoicedLpcOrder],
617 &audio_history[temp_index + kUnvoicedLpcOrder],
618 sizeof(int16_t) * fs_mult_lpc_analysis_len);
619 CrossCorrelationWithAutoShift(
620 &temp_signal[kUnvoicedLpcOrder], &temp_signal[kUnvoicedLpcOrder],
621 fs_mult_lpc_analysis_len, kUnvoicedLpcOrder + 1, -1, auto_correlation);
622 delete [] temp_signal;
623
624 // Verify that variance is positive.
625 if (auto_correlation[0] > 0) {
626 // Estimate AR filter parameters using Levinson-Durbin algorithm;
627 // kUnvoicedLpcOrder + 1 filter coefficients.
628 int16_t stability = WebRtcSpl_LevinsonDurbin(auto_correlation,
629 parameters.ar_filter,
630 reflection_coeff,
631 kUnvoicedLpcOrder);
632
633 // Keep filter parameters only if filter is stable.
634 if (stability != 1) {
635 // Set first coefficient to 4096 (1.0 in Q12).
636 parameters.ar_filter[0] = 4096;
637 // Set remaining |kUnvoicedLpcOrder| coefficients to zero.
638 WebRtcSpl_MemSetW16(parameters.ar_filter + 1, 0, kUnvoicedLpcOrder);
639 }
640 }
641
642 if (channel_ix == 0) {
643 // Extract a noise segment.
644 size_t noise_length;
645 if (distortion_lag < 40) {
646 noise_length = 2 * distortion_lag + 30;
647 } else {
648 noise_length = distortion_lag + 30;
649 }
650 if (noise_length <= RandomVector::kRandomTableSize) {
651 memcpy(random_vector, RandomVector::kRandomTable,
652 sizeof(int16_t) * noise_length);
653 } else {
654 // Only applies to SWB where length could be larger than
655 // |kRandomTableSize|.
656 memcpy(random_vector, RandomVector::kRandomTable,
657 sizeof(int16_t) * RandomVector::kRandomTableSize);
658 assert(noise_length <= kMaxSampleRate / 8000 * 120 + 30);
659 random_vector_->IncreaseSeedIncrement(2);
660 random_vector_->Generate(
661 noise_length - RandomVector::kRandomTableSize,
662 &random_vector[RandomVector::kRandomTableSize]);
663 }
664 }
665
666 // Set up state vector and calculate scale factor for unvoiced filtering.
667 memcpy(parameters.ar_filter_state,
668 &(audio_history[signal_length - kUnvoicedLpcOrder]),
669 sizeof(int16_t) * kUnvoicedLpcOrder);
670 memcpy(unvoiced_vector - kUnvoicedLpcOrder,
671 &(audio_history[signal_length - 128 - kUnvoicedLpcOrder]),
672 sizeof(int16_t) * kUnvoicedLpcOrder);
673 WebRtcSpl_FilterMAFastQ12(&audio_history[signal_length - 128],
674 unvoiced_vector,
675 parameters.ar_filter,
676 kUnvoicedLpcOrder + 1,
677 128);
678 const int unvoiced_max_abs = [&] {
679 const int16_t max_abs = WebRtcSpl_MaxAbsValueW16(unvoiced_vector, 128);
680 // Since WebRtcSpl_MaxAbsValueW16 returns 2^15 - 1 when the input contains
681 // -2^15, we have to conservatively bump the return value by 1
682 // if it is 2^15 - 1.
683 return max_abs == WEBRTC_SPL_WORD16_MAX ? max_abs + 1 : max_abs;
684 }();
685 // Pick the smallest n such that 2^n > unvoiced_max_abs; then the maximum
686 // value of the dot product is less than 2^7 * 2^(2*n) = 2^(2*n + 7), so to
687 // prevent overflows we want 2n + 7 <= 31, which means we should shift by
688 // 2n + 7 - 31 bits, if this value is greater than zero.
689 int unvoiced_prescale =
690 std::max(0, 2 * WebRtcSpl_GetSizeInBits(unvoiced_max_abs) - 24);
691
692 int32_t unvoiced_energy = WebRtcSpl_DotProductWithScale(unvoiced_vector,
693 unvoiced_vector,
694 128,
695 unvoiced_prescale);
696
697 // Normalize |unvoiced_energy| to 28 or 29 bits to preserve sqrt() accuracy.
698 int16_t unvoiced_scale = WebRtcSpl_NormW32(unvoiced_energy) - 3;
699 // Make sure we do an odd number of shifts since we already have 7 shifts
700 // from dividing with 128 earlier. This will make the total scale factor
701 // even, which is suitable for the sqrt.
702 unvoiced_scale += ((unvoiced_scale & 0x1) ^ 0x1);
703 unvoiced_energy = WEBRTC_SPL_SHIFT_W32(unvoiced_energy, unvoiced_scale);
704 int16_t unvoiced_gain =
705 static_cast<int16_t>(WebRtcSpl_SqrtFloor(unvoiced_energy));
706 parameters.ar_gain_scale = 13
707 + (unvoiced_scale + 7 - unvoiced_prescale) / 2;
708 parameters.ar_gain = unvoiced_gain;
709
710 // Calculate voice_mix_factor from corr_coefficient.
711 // Let x = corr_coefficient. Then, we compute:
712 // if (x > 0.48)
713 // voice_mix_factor = (-5179 + 19931x - 16422x^2 + 5776x^3) / 4096;
714 // else
715 // voice_mix_factor = 0;
716 if (corr_coefficient > 7875) {
717 int16_t x1, x2, x3;
718 // |corr_coefficient| is in Q14.
719 x1 = static_cast<int16_t>(corr_coefficient);
720 x2 = (x1 * x1) >> 14; // Shift 14 to keep result in Q14.
721 x3 = (x1 * x2) >> 14;
722 static const int kCoefficients[4] = { -5179, 19931, -16422, 5776 };
723 int32_t temp_sum = kCoefficients[0] * 16384;
724 temp_sum += kCoefficients[1] * x1;
725 temp_sum += kCoefficients[2] * x2;
726 temp_sum += kCoefficients[3] * x3;
727 parameters.voice_mix_factor =
728 static_cast<int16_t>(std::min(temp_sum / 4096, 16384));
729 parameters.voice_mix_factor = std::max(parameters.voice_mix_factor,
730 static_cast<int16_t>(0));
731 } else {
732 parameters.voice_mix_factor = 0;
733 }
734
735 // Calculate muting slope. Reuse value from earlier scaling of
736 // |expand_vector0| and |expand_vector1|.
737 int16_t slope = amplitude_ratio;
738 if (slope > 12288) {
739 // slope > 1.5.
740 // Calculate (1 - (1 / slope)) / distortion_lag =
741 // (slope - 1) / (distortion_lag * slope).
742 // |slope| is in Q13, so 1 corresponds to 8192. Shift up to Q25 before
743 // the division.
744 // Shift the denominator from Q13 to Q5 before the division. The result of
745 // the division will then be in Q20.
746 int temp_ratio = WebRtcSpl_DivW32W16(
747 (slope - 8192) << 12,
748 static_cast<int16_t>((distortion_lag * slope) >> 8));
749 if (slope > 14746) {
750 // slope > 1.8.
751 // Divide by 2, with proper rounding.
752 parameters.mute_slope = (temp_ratio + 1) / 2;
753 } else {
754 // Divide by 8, with proper rounding.
755 parameters.mute_slope = (temp_ratio + 4) / 8;
756 }
757 parameters.onset = true;
758 } else {
759 // Calculate (1 - slope) / distortion_lag.
760 // Shift |slope| by 7 to Q20 before the division. The result is in Q20.
761 parameters.mute_slope = WebRtcSpl_DivW32W16(
762 (8192 - slope) * 128, static_cast<int16_t>(distortion_lag));
763 if (parameters.voice_mix_factor <= 13107) {
764 // Make sure the mute factor decreases from 1.0 to 0.9 in no more than
765 // 6.25 ms.
766 // mute_slope >= 0.005 / fs_mult in Q20.
767 parameters.mute_slope = std::max(5243 / fs_mult, parameters.mute_slope);
768 } else if (slope > 8028) {
769 parameters.mute_slope = 0;
770 }
771 parameters.onset = false;
772 }
773 }
774 }
775
ChannelParameters()776 Expand::ChannelParameters::ChannelParameters()
777 : mute_factor(16384),
778 ar_gain(0),
779 ar_gain_scale(0),
780 voice_mix_factor(0),
781 current_voice_mix_factor(0),
782 onset(false),
783 mute_slope(0) {
784 memset(ar_filter, 0, sizeof(ar_filter));
785 memset(ar_filter_state, 0, sizeof(ar_filter_state));
786 }
787
Correlation(const int16_t * input,size_t input_length,int16_t * output) const788 void Expand::Correlation(const int16_t* input,
789 size_t input_length,
790 int16_t* output) const {
791 // Set parameters depending on sample rate.
792 const int16_t* filter_coefficients;
793 size_t num_coefficients;
794 int16_t downsampling_factor;
795 if (fs_hz_ == 8000) {
796 num_coefficients = 3;
797 downsampling_factor = 2;
798 filter_coefficients = DspHelper::kDownsample8kHzTbl;
799 } else if (fs_hz_ == 16000) {
800 num_coefficients = 5;
801 downsampling_factor = 4;
802 filter_coefficients = DspHelper::kDownsample16kHzTbl;
803 } else if (fs_hz_ == 32000) {
804 num_coefficients = 7;
805 downsampling_factor = 8;
806 filter_coefficients = DspHelper::kDownsample32kHzTbl;
807 } else { // fs_hz_ == 48000.
808 num_coefficients = 7;
809 downsampling_factor = 12;
810 filter_coefficients = DspHelper::kDownsample48kHzTbl;
811 }
812
813 // Correlate from lag 10 to lag 60 in downsampled domain.
814 // (Corresponds to 20-120 for narrow-band, 40-240 for wide-band, and so on.)
815 static const size_t kCorrelationStartLag = 10;
816 static const size_t kNumCorrelationLags = 54;
817 static const size_t kCorrelationLength = 60;
818 // Downsample to 4 kHz sample rate.
819 static const size_t kDownsampledLength = kCorrelationStartLag
820 + kNumCorrelationLags + kCorrelationLength;
821 int16_t downsampled_input[kDownsampledLength];
822 static const size_t kFilterDelay = 0;
823 WebRtcSpl_DownsampleFast(
824 input + input_length - kDownsampledLength * downsampling_factor,
825 kDownsampledLength * downsampling_factor, downsampled_input,
826 kDownsampledLength, filter_coefficients, num_coefficients,
827 downsampling_factor, kFilterDelay);
828
829 // Normalize |downsampled_input| to using all 16 bits.
830 int16_t max_value = WebRtcSpl_MaxAbsValueW16(downsampled_input,
831 kDownsampledLength);
832 int16_t norm_shift = 16 - WebRtcSpl_NormW32(max_value);
833 WebRtcSpl_VectorBitShiftW16(downsampled_input, kDownsampledLength,
834 downsampled_input, norm_shift);
835
836 int32_t correlation[kNumCorrelationLags];
837 CrossCorrelationWithAutoShift(
838 &downsampled_input[kDownsampledLength - kCorrelationLength],
839 &downsampled_input[kDownsampledLength - kCorrelationLength
840 - kCorrelationStartLag],
841 kCorrelationLength, kNumCorrelationLags, -1, correlation);
842
843 // Normalize and move data from 32-bit to 16-bit vector.
844 int32_t max_correlation = WebRtcSpl_MaxAbsValueW32(correlation,
845 kNumCorrelationLags);
846 int16_t norm_shift2 = static_cast<int16_t>(
847 std::max(18 - WebRtcSpl_NormW32(max_correlation), 0));
848 WebRtcSpl_VectorBitShiftW32ToW16(output, kNumCorrelationLags, correlation,
849 norm_shift2);
850 }
851
UpdateLagIndex()852 void Expand::UpdateLagIndex() {
853 current_lag_index_ = current_lag_index_ + lag_index_direction_;
854 // Change direction if needed.
855 if (current_lag_index_ <= 0) {
856 lag_index_direction_ = 1;
857 }
858 if (current_lag_index_ >= kNumLags - 1) {
859 lag_index_direction_ = -1;
860 }
861 }
862
Create(BackgroundNoise * background_noise,SyncBuffer * sync_buffer,RandomVector * random_vector,StatisticsCalculator * statistics,int fs,size_t num_channels) const863 Expand* ExpandFactory::Create(BackgroundNoise* background_noise,
864 SyncBuffer* sync_buffer,
865 RandomVector* random_vector,
866 StatisticsCalculator* statistics,
867 int fs,
868 size_t num_channels) const {
869 return new Expand(background_noise, sync_buffer, random_vector, statistics,
870 fs, num_channels);
871 }
872
873 // TODO(turajs): This can be moved to BackgroundNoise class.
GenerateBackgroundNoise(int16_t * random_vector,size_t channel,int mute_slope,bool too_many_expands,size_t num_noise_samples,int16_t * buffer)874 void Expand::GenerateBackgroundNoise(int16_t* random_vector,
875 size_t channel,
876 int mute_slope,
877 bool too_many_expands,
878 size_t num_noise_samples,
879 int16_t* buffer) {
880 static const size_t kNoiseLpcOrder = BackgroundNoise::kMaxLpcOrder;
881 int16_t scaled_random_vector[kMaxSampleRate / 8000 * 125];
882 assert(num_noise_samples <= (kMaxSampleRate / 8000 * 125));
883 int16_t* noise_samples = &buffer[kNoiseLpcOrder];
884 if (background_noise_->initialized()) {
885 // Use background noise parameters.
886 memcpy(noise_samples - kNoiseLpcOrder,
887 background_noise_->FilterState(channel),
888 sizeof(int16_t) * kNoiseLpcOrder);
889
890 int dc_offset = 0;
891 if (background_noise_->ScaleShift(channel) > 1) {
892 dc_offset = 1 << (background_noise_->ScaleShift(channel) - 1);
893 }
894
895 // Scale random vector to correct energy level.
896 WebRtcSpl_AffineTransformVector(
897 scaled_random_vector, random_vector,
898 background_noise_->Scale(channel), dc_offset,
899 background_noise_->ScaleShift(channel),
900 num_noise_samples);
901
902 WebRtcSpl_FilterARFastQ12(scaled_random_vector, noise_samples,
903 background_noise_->Filter(channel),
904 kNoiseLpcOrder + 1,
905 num_noise_samples);
906
907 background_noise_->SetFilterState(
908 channel,
909 &(noise_samples[num_noise_samples - kNoiseLpcOrder]),
910 kNoiseLpcOrder);
911
912 // Unmute the background noise.
913 int16_t bgn_mute_factor = background_noise_->MuteFactor(channel);
914 NetEq::BackgroundNoiseMode bgn_mode = background_noise_->mode();
915 if (bgn_mode == NetEq::kBgnFade && too_many_expands &&
916 bgn_mute_factor > 0) {
917 // Fade BGN to zero.
918 // Calculate muting slope, approximately -2^18 / fs_hz.
919 int mute_slope;
920 if (fs_hz_ == 8000) {
921 mute_slope = -32;
922 } else if (fs_hz_ == 16000) {
923 mute_slope = -16;
924 } else if (fs_hz_ == 32000) {
925 mute_slope = -8;
926 } else {
927 mute_slope = -5;
928 }
929 // Use UnmuteSignal function with negative slope.
930 // |bgn_mute_factor| is in Q14. |mute_slope| is in Q20.
931 DspHelper::UnmuteSignal(noise_samples,
932 num_noise_samples,
933 &bgn_mute_factor,
934 mute_slope,
935 noise_samples);
936 } else if (bgn_mute_factor < 16384) {
937 // If mode is kBgnOn, or if kBgnFade has started fading,
938 // use regular |mute_slope|.
939 if (!stop_muting_ && bgn_mode != NetEq::kBgnOff &&
940 !(bgn_mode == NetEq::kBgnFade && too_many_expands)) {
941 DspHelper::UnmuteSignal(noise_samples,
942 static_cast<int>(num_noise_samples),
943 &bgn_mute_factor,
944 mute_slope,
945 noise_samples);
946 } else {
947 // kBgnOn and stop muting, or
948 // kBgnOff (mute factor is always 0), or
949 // kBgnFade has reached 0.
950 WebRtcSpl_AffineTransformVector(noise_samples, noise_samples,
951 bgn_mute_factor, 8192, 14,
952 num_noise_samples);
953 }
954 }
955 // Update mute_factor in BackgroundNoise class.
956 background_noise_->SetMuteFactor(channel, bgn_mute_factor);
957 } else {
958 // BGN parameters have not been initialized; use zero noise.
959 memset(noise_samples, 0, sizeof(int16_t) * num_noise_samples);
960 }
961 }
962
GenerateRandomVector(int16_t seed_increment,size_t length,int16_t * random_vector)963 void Expand::GenerateRandomVector(int16_t seed_increment,
964 size_t length,
965 int16_t* random_vector) {
966 // TODO(turajs): According to hlundin The loop should not be needed. Should be
967 // just as good to generate all of the vector in one call.
968 size_t samples_generated = 0;
969 const size_t kMaxRandSamples = RandomVector::kRandomTableSize;
970 while (samples_generated < length) {
971 size_t rand_length = std::min(length - samples_generated, kMaxRandSamples);
972 random_vector_->IncreaseSeedIncrement(seed_increment);
973 random_vector_->Generate(rand_length, &random_vector[samples_generated]);
974 samples_generated += rand_length;
975 }
976 }
977
978 } // namespace webrtc
979