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/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 "webrtc/base/safe_conversions.h"
20 #include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
21 #include "webrtc/modules/audio_coding/neteq/background_noise.h"
22 #include "webrtc/modules/audio_coding/neteq/dsp_helper.h"
23 #include "webrtc/modules/audio_coding/neteq/random_vector.h"
24 #include "webrtc/modules/audio_coding/neteq/statistics_calculator.h"
25 #include "webrtc/modules/audio_coding/neteq/sync_buffer.h"
26
27 namespace webrtc {
28
Expand(BackgroundNoise * background_noise,SyncBuffer * sync_buffer,RandomVector * random_vector,StatisticsCalculator * statistics,int fs,size_t num_channels)29 Expand::Expand(BackgroundNoise* background_noise,
30 SyncBuffer* sync_buffer,
31 RandomVector* random_vector,
32 StatisticsCalculator* statistics,
33 int fs,
34 size_t num_channels)
35 : random_vector_(random_vector),
36 sync_buffer_(sync_buffer),
37 first_expand_(true),
38 fs_hz_(fs),
39 num_channels_(num_channels),
40 consecutive_expands_(0),
41 background_noise_(background_noise),
42 statistics_(statistics),
43 overlap_length_(5 * fs / 8000),
44 lag_index_direction_(0),
45 current_lag_index_(0),
46 stop_muting_(false),
47 expand_duration_samples_(0),
48 channel_parameters_(new ChannelParameters[num_channels_]) {
49 assert(fs == 8000 || fs == 16000 || fs == 32000 || fs == 48000);
50 assert(fs <= static_cast<int>(kMaxSampleRate)); // Should not be possible.
51 assert(num_channels_ > 0);
52 memset(expand_lags_, 0, sizeof(expand_lags_));
53 Reset();
54 }
55
56 Expand::~Expand() = default;
57
Reset()58 void Expand::Reset() {
59 first_expand_ = true;
60 consecutive_expands_ = 0;
61 max_lag_ = 0;
62 for (size_t ix = 0; ix < num_channels_; ++ix) {
63 channel_parameters_[ix].expand_vector0.Clear();
64 channel_parameters_[ix].expand_vector1.Clear();
65 }
66 }
67
Process(AudioMultiVector * output)68 int Expand::Process(AudioMultiVector* output) {
69 int16_t random_vector[kMaxSampleRate / 8000 * 120 + 30];
70 int16_t scaled_random_vector[kMaxSampleRate / 8000 * 125];
71 static const int kTempDataSize = 3600;
72 int16_t temp_data[kTempDataSize]; // TODO(hlundin) Remove this.
73 int16_t* voiced_vector_storage = temp_data;
74 int16_t* voiced_vector = &voiced_vector_storage[overlap_length_];
75 static const size_t kNoiseLpcOrder = BackgroundNoise::kMaxLpcOrder;
76 int16_t unvoiced_array_memory[kNoiseLpcOrder + kMaxSampleRate / 8000 * 125];
77 int16_t* unvoiced_vector = unvoiced_array_memory + kUnvoicedLpcOrder;
78 int16_t* noise_vector = unvoiced_array_memory + kNoiseLpcOrder;
79
80 int fs_mult = fs_hz_ / 8000;
81
82 if (first_expand_) {
83 // Perform initial setup if this is the first expansion since last reset.
84 AnalyzeSignal(random_vector);
85 first_expand_ = false;
86 expand_duration_samples_ = 0;
87 } else {
88 // This is not the first expansion, parameters are already estimated.
89 // Extract a noise segment.
90 size_t rand_length = max_lag_;
91 // This only applies to SWB where length could be larger than 256.
92 assert(rand_length <= kMaxSampleRate / 8000 * 120 + 30);
93 GenerateRandomVector(2, rand_length, random_vector);
94 }
95
96
97 // Generate signal.
98 UpdateLagIndex();
99
100 // Voiced part.
101 // Generate a weighted vector with the current lag.
102 size_t expansion_vector_length = max_lag_ + overlap_length_;
103 size_t current_lag = expand_lags_[current_lag_index_];
104 // Copy lag+overlap data.
105 size_t expansion_vector_position = expansion_vector_length - current_lag -
106 overlap_length_;
107 size_t temp_length = current_lag + overlap_length_;
108 for (size_t channel_ix = 0; channel_ix < num_channels_; ++channel_ix) {
109 ChannelParameters& parameters = channel_parameters_[channel_ix];
110 if (current_lag_index_ == 0) {
111 // Use only expand_vector0.
112 assert(expansion_vector_position + temp_length <=
113 parameters.expand_vector0.Size());
114 memcpy(voiced_vector_storage,
115 ¶meters.expand_vector0[expansion_vector_position],
116 sizeof(int16_t) * temp_length);
117 } else if (current_lag_index_ == 1) {
118 // Mix 3/4 of expand_vector0 with 1/4 of expand_vector1.
119 WebRtcSpl_ScaleAndAddVectorsWithRound(
120 ¶meters.expand_vector0[expansion_vector_position], 3,
121 ¶meters.expand_vector1[expansion_vector_position], 1, 2,
122 voiced_vector_storage, temp_length);
123 } else if (current_lag_index_ == 2) {
124 // Mix 1/2 of expand_vector0 with 1/2 of expand_vector1.
125 assert(expansion_vector_position + temp_length <=
126 parameters.expand_vector0.Size());
127 assert(expansion_vector_position + temp_length <=
128 parameters.expand_vector1.Size());
129 WebRtcSpl_ScaleAndAddVectorsWithRound(
130 ¶meters.expand_vector0[expansion_vector_position], 1,
131 ¶meters.expand_vector1[expansion_vector_position], 1, 1,
132 voiced_vector_storage, temp_length);
133 }
134
135 // Get tapering window parameters. Values are in Q15.
136 int16_t muting_window, muting_window_increment;
137 int16_t unmuting_window, unmuting_window_increment;
138 if (fs_hz_ == 8000) {
139 muting_window = DspHelper::kMuteFactorStart8kHz;
140 muting_window_increment = DspHelper::kMuteFactorIncrement8kHz;
141 unmuting_window = DspHelper::kUnmuteFactorStart8kHz;
142 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement8kHz;
143 } else if (fs_hz_ == 16000) {
144 muting_window = DspHelper::kMuteFactorStart16kHz;
145 muting_window_increment = DspHelper::kMuteFactorIncrement16kHz;
146 unmuting_window = DspHelper::kUnmuteFactorStart16kHz;
147 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement16kHz;
148 } else if (fs_hz_ == 32000) {
149 muting_window = DspHelper::kMuteFactorStart32kHz;
150 muting_window_increment = DspHelper::kMuteFactorIncrement32kHz;
151 unmuting_window = DspHelper::kUnmuteFactorStart32kHz;
152 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement32kHz;
153 } else { // fs_ == 48000
154 muting_window = DspHelper::kMuteFactorStart48kHz;
155 muting_window_increment = DspHelper::kMuteFactorIncrement48kHz;
156 unmuting_window = DspHelper::kUnmuteFactorStart48kHz;
157 unmuting_window_increment = DspHelper::kUnmuteFactorIncrement48kHz;
158 }
159
160 // Smooth the expanded if it has not been muted to a low amplitude and
161 // |current_voice_mix_factor| is larger than 0.5.
162 if ((parameters.mute_factor > 819) &&
163 (parameters.current_voice_mix_factor > 8192)) {
164 size_t start_ix = sync_buffer_->Size() - overlap_length_;
165 for (size_t i = 0; i < overlap_length_; i++) {
166 // Do overlap add between new vector and overlap.
167 (*sync_buffer_)[channel_ix][start_ix + i] =
168 (((*sync_buffer_)[channel_ix][start_ix + i] * muting_window) +
169 (((parameters.mute_factor * voiced_vector_storage[i]) >> 14) *
170 unmuting_window) + 16384) >> 15;
171 muting_window += muting_window_increment;
172 unmuting_window += unmuting_window_increment;
173 }
174 } else if (parameters.mute_factor == 0) {
175 // The expanded signal will consist of only comfort noise if
176 // mute_factor = 0. Set the output length to 15 ms for best noise
177 // production.
178 // TODO(hlundin): This has been disabled since the length of
179 // parameters.expand_vector0 and parameters.expand_vector1 no longer
180 // match with expand_lags_, causing invalid reads and writes. Is it a good
181 // idea to enable this again, and solve the vector size problem?
182 // max_lag_ = fs_mult * 120;
183 // expand_lags_[0] = fs_mult * 120;
184 // expand_lags_[1] = fs_mult * 120;
185 // expand_lags_[2] = fs_mult * 120;
186 }
187
188 // Unvoiced part.
189 // Filter |scaled_random_vector| through |ar_filter_|.
190 memcpy(unvoiced_vector - kUnvoicedLpcOrder, parameters.ar_filter_state,
191 sizeof(int16_t) * kUnvoicedLpcOrder);
192 int32_t add_constant = 0;
193 if (parameters.ar_gain_scale > 0) {
194 add_constant = 1 << (parameters.ar_gain_scale - 1);
195 }
196 WebRtcSpl_AffineTransformVector(scaled_random_vector, random_vector,
197 parameters.ar_gain, add_constant,
198 parameters.ar_gain_scale,
199 current_lag);
200 WebRtcSpl_FilterARFastQ12(scaled_random_vector, unvoiced_vector,
201 parameters.ar_filter, kUnvoicedLpcOrder + 1,
202 current_lag);
203 memcpy(parameters.ar_filter_state,
204 &(unvoiced_vector[current_lag - kUnvoicedLpcOrder]),
205 sizeof(int16_t) * kUnvoicedLpcOrder);
206
207 // Combine voiced and unvoiced contributions.
208
209 // Set a suitable cross-fading slope.
210 // For lag =
211 // <= 31 * fs_mult => go from 1 to 0 in about 8 ms;
212 // (>= 31 .. <= 63) * fs_mult => go from 1 to 0 in about 16 ms;
213 // >= 64 * fs_mult => go from 1 to 0 in about 32 ms.
214 // temp_shift = getbits(max_lag_) - 5.
215 int temp_shift =
216 (31 - WebRtcSpl_NormW32(rtc::checked_cast<int32_t>(max_lag_))) - 5;
217 int16_t mix_factor_increment = 256 >> temp_shift;
218 if (stop_muting_) {
219 mix_factor_increment = 0;
220 }
221
222 // Create combined signal by shifting in more and more of unvoiced part.
223 temp_shift = 8 - temp_shift; // = getbits(mix_factor_increment).
224 size_t temp_length = (parameters.current_voice_mix_factor -
225 parameters.voice_mix_factor) >> temp_shift;
226 temp_length = std::min(temp_length, current_lag);
227 DspHelper::CrossFade(voiced_vector, unvoiced_vector, temp_length,
228 ¶meters.current_voice_mix_factor,
229 mix_factor_increment, temp_data);
230
231 // End of cross-fading period was reached before end of expanded signal
232 // path. Mix the rest with a fixed mixing factor.
233 if (temp_length < current_lag) {
234 if (mix_factor_increment != 0) {
235 parameters.current_voice_mix_factor = parameters.voice_mix_factor;
236 }
237 int16_t temp_scale = 16384 - parameters.current_voice_mix_factor;
238 WebRtcSpl_ScaleAndAddVectorsWithRound(
239 voiced_vector + temp_length, parameters.current_voice_mix_factor,
240 unvoiced_vector + temp_length, temp_scale, 14,
241 temp_data + temp_length, current_lag - temp_length);
242 }
243
244 // Select muting slope depending on how many consecutive expands we have
245 // done.
246 if (consecutive_expands_ == 3) {
247 // Let the mute factor decrease from 1.0 to 0.95 in 6.25 ms.
248 // mute_slope = 0.0010 / fs_mult in Q20.
249 parameters.mute_slope = std::max(parameters.mute_slope, 1049 / fs_mult);
250 }
251 if (consecutive_expands_ == 7) {
252 // Let the mute factor decrease from 1.0 to 0.90 in 6.25 ms.
253 // mute_slope = 0.0020 / fs_mult in Q20.
254 parameters.mute_slope = std::max(parameters.mute_slope, 2097 / fs_mult);
255 }
256
257 // Mute segment according to slope value.
258 if ((consecutive_expands_ != 0) || !parameters.onset) {
259 // Mute to the previous level, then continue with the muting.
260 WebRtcSpl_AffineTransformVector(temp_data, temp_data,
261 parameters.mute_factor, 8192,
262 14, current_lag);
263
264 if (!stop_muting_) {
265 DspHelper::MuteSignal(temp_data, parameters.mute_slope, current_lag);
266
267 // Shift by 6 to go from Q20 to Q14.
268 // TODO(hlundin): Adding 8192 before shifting 6 steps seems wrong.
269 // Legacy.
270 int16_t gain = static_cast<int16_t>(16384 -
271 (((current_lag * parameters.mute_slope) + 8192) >> 6));
272 gain = ((gain * parameters.mute_factor) + 8192) >> 14;
273
274 // Guard against getting stuck with very small (but sometimes audible)
275 // gain.
276 if ((consecutive_expands_ > 3) && (gain >= parameters.mute_factor)) {
277 parameters.mute_factor = 0;
278 } else {
279 parameters.mute_factor = gain;
280 }
281 }
282 }
283
284 // Background noise part.
285 GenerateBackgroundNoise(random_vector,
286 channel_ix,
287 channel_parameters_[channel_ix].mute_slope,
288 TooManyExpands(),
289 current_lag,
290 unvoiced_array_memory);
291
292 // Add background noise to the combined voiced-unvoiced signal.
293 for (size_t i = 0; i < current_lag; i++) {
294 temp_data[i] = temp_data[i] + noise_vector[i];
295 }
296 if (channel_ix == 0) {
297 output->AssertSize(current_lag);
298 } else {
299 assert(output->Size() == current_lag);
300 }
301 memcpy(&(*output)[channel_ix][0], temp_data,
302 sizeof(temp_data[0]) * current_lag);
303 }
304
305 // Increase call number and cap it.
306 consecutive_expands_ = consecutive_expands_ >= kMaxConsecutiveExpands ?
307 kMaxConsecutiveExpands : consecutive_expands_ + 1;
308 expand_duration_samples_ += output->Size();
309 // Clamp the duration counter at 2 seconds.
310 expand_duration_samples_ =
311 std::min(expand_duration_samples_, rtc::checked_cast<size_t>(fs_hz_ * 2));
312 return 0;
313 }
314
SetParametersForNormalAfterExpand()315 void Expand::SetParametersForNormalAfterExpand() {
316 current_lag_index_ = 0;
317 lag_index_direction_ = 0;
318 stop_muting_ = true; // Do not mute signal any more.
319 statistics_->LogDelayedPacketOutageEvent(
320 rtc::checked_cast<int>(expand_duration_samples_) / (fs_hz_ / 1000));
321 }
322
SetParametersForMergeAfterExpand()323 void Expand::SetParametersForMergeAfterExpand() {
324 current_lag_index_ = -1; /* out of the 3 possible ones */
325 lag_index_direction_ = 1; /* make sure we get the "optimal" lag */
326 stop_muting_ = true;
327 }
328
overlap_length() const329 size_t Expand::overlap_length() const {
330 return overlap_length_;
331 }
332
InitializeForAnExpandPeriod()333 void Expand::InitializeForAnExpandPeriod() {
334 lag_index_direction_ = 1;
335 current_lag_index_ = -1;
336 stop_muting_ = false;
337 random_vector_->set_seed_increment(1);
338 consecutive_expands_ = 0;
339 for (size_t ix = 0; ix < num_channels_; ++ix) {
340 channel_parameters_[ix].current_voice_mix_factor = 16384; // 1.0 in Q14.
341 channel_parameters_[ix].mute_factor = 16384; // 1.0 in Q14.
342 // Start with 0 gain for background noise.
343 background_noise_->SetMuteFactor(ix, 0);
344 }
345 }
346
TooManyExpands()347 bool Expand::TooManyExpands() {
348 return consecutive_expands_ >= kMaxConsecutiveExpands;
349 }
350
AnalyzeSignal(int16_t * random_vector)351 void Expand::AnalyzeSignal(int16_t* random_vector) {
352 int32_t auto_correlation[kUnvoicedLpcOrder + 1];
353 int16_t reflection_coeff[kUnvoicedLpcOrder];
354 int16_t correlation_vector[kMaxSampleRate / 8000 * 102];
355 size_t best_correlation_index[kNumCorrelationCandidates];
356 int16_t best_correlation[kNumCorrelationCandidates];
357 size_t best_distortion_index[kNumCorrelationCandidates];
358 int16_t best_distortion[kNumCorrelationCandidates];
359 int32_t correlation_vector2[(99 * kMaxSampleRate / 8000) + 1];
360 int32_t best_distortion_w32[kNumCorrelationCandidates];
361 static const size_t kNoiseLpcOrder = BackgroundNoise::kMaxLpcOrder;
362 int16_t unvoiced_array_memory[kNoiseLpcOrder + kMaxSampleRate / 8000 * 125];
363 int16_t* unvoiced_vector = unvoiced_array_memory + kUnvoicedLpcOrder;
364
365 int fs_mult = fs_hz_ / 8000;
366
367 // Pre-calculate common multiplications with fs_mult.
368 size_t fs_mult_4 = static_cast<size_t>(fs_mult * 4);
369 size_t fs_mult_20 = static_cast<size_t>(fs_mult * 20);
370 size_t fs_mult_120 = static_cast<size_t>(fs_mult * 120);
371 size_t fs_mult_dist_len = fs_mult * kDistortionLength;
372 size_t fs_mult_lpc_analysis_len = fs_mult * kLpcAnalysisLength;
373
374 const size_t signal_length = static_cast<size_t>(256 * fs_mult);
375 const int16_t* audio_history =
376 &(*sync_buffer_)[0][sync_buffer_->Size() - signal_length];
377
378 // Initialize.
379 InitializeForAnExpandPeriod();
380
381 // Calculate correlation in downsampled domain (4 kHz sample rate).
382 int correlation_scale;
383 size_t correlation_length = 51; // TODO(hlundin): Legacy bit-exactness.
384 // If it is decided to break bit-exactness |correlation_length| should be
385 // initialized to the return value of Correlation().
386 Correlation(audio_history, signal_length, correlation_vector,
387 &correlation_scale);
388
389 // Find peaks in correlation vector.
390 DspHelper::PeakDetection(correlation_vector, correlation_length,
391 kNumCorrelationCandidates, fs_mult,
392 best_correlation_index, best_correlation);
393
394 // Adjust peak locations; cross-correlation lags start at 2.5 ms
395 // (20 * fs_mult samples).
396 best_correlation_index[0] += fs_mult_20;
397 best_correlation_index[1] += fs_mult_20;
398 best_correlation_index[2] += fs_mult_20;
399
400 // Calculate distortion around the |kNumCorrelationCandidates| best lags.
401 int distortion_scale = 0;
402 for (size_t i = 0; i < kNumCorrelationCandidates; i++) {
403 size_t min_index = std::max(fs_mult_20,
404 best_correlation_index[i] - fs_mult_4);
405 size_t max_index = std::min(fs_mult_120 - 1,
406 best_correlation_index[i] + fs_mult_4);
407 best_distortion_index[i] = DspHelper::MinDistortion(
408 &(audio_history[signal_length - fs_mult_dist_len]), min_index,
409 max_index, fs_mult_dist_len, &best_distortion_w32[i]);
410 distortion_scale = std::max(16 - WebRtcSpl_NormW32(best_distortion_w32[i]),
411 distortion_scale);
412 }
413 // Shift the distortion values to fit in 16 bits.
414 WebRtcSpl_VectorBitShiftW32ToW16(best_distortion, kNumCorrelationCandidates,
415 best_distortion_w32, distortion_scale);
416
417 // Find the maximizing index |i| of the cost function
418 // f[i] = best_correlation[i] / best_distortion[i].
419 int32_t best_ratio = std::numeric_limits<int32_t>::min();
420 size_t best_index = std::numeric_limits<size_t>::max();
421 for (size_t i = 0; i < kNumCorrelationCandidates; ++i) {
422 int32_t ratio;
423 if (best_distortion[i] > 0) {
424 ratio = (best_correlation[i] << 16) / best_distortion[i];
425 } else if (best_correlation[i] == 0) {
426 ratio = 0; // No correlation set result to zero.
427 } else {
428 ratio = std::numeric_limits<int32_t>::max(); // Denominator is zero.
429 }
430 if (ratio > best_ratio) {
431 best_index = i;
432 best_ratio = ratio;
433 }
434 }
435
436 size_t distortion_lag = best_distortion_index[best_index];
437 size_t correlation_lag = best_correlation_index[best_index];
438 max_lag_ = std::max(distortion_lag, correlation_lag);
439
440 // Calculate the exact best correlation in the range between
441 // |correlation_lag| and |distortion_lag|.
442 correlation_length =
443 std::max(std::min(distortion_lag + 10, fs_mult_120),
444 static_cast<size_t>(60 * fs_mult));
445
446 size_t start_index = std::min(distortion_lag, correlation_lag);
447 size_t correlation_lags = static_cast<size_t>(
448 WEBRTC_SPL_ABS_W16((distortion_lag-correlation_lag)) + 1);
449 assert(correlation_lags <= static_cast<size_t>(99 * fs_mult + 1));
450
451 for (size_t channel_ix = 0; channel_ix < num_channels_; ++channel_ix) {
452 ChannelParameters& parameters = channel_parameters_[channel_ix];
453 // Calculate suitable scaling.
454 int16_t signal_max = WebRtcSpl_MaxAbsValueW16(
455 &audio_history[signal_length - correlation_length - start_index
456 - correlation_lags],
457 correlation_length + start_index + correlation_lags - 1);
458 correlation_scale = (31 - WebRtcSpl_NormW32(signal_max * signal_max)) +
459 (31 - WebRtcSpl_NormW32(static_cast<int32_t>(correlation_length))) - 31;
460 correlation_scale = std::max(0, correlation_scale);
461
462 // Calculate the correlation, store in |correlation_vector2|.
463 WebRtcSpl_CrossCorrelation(
464 correlation_vector2,
465 &(audio_history[signal_length - correlation_length]),
466 &(audio_history[signal_length - correlation_length - start_index]),
467 correlation_length, correlation_lags, correlation_scale, -1);
468
469 // Find maximizing index.
470 best_index = WebRtcSpl_MaxIndexW32(correlation_vector2, correlation_lags);
471 int32_t max_correlation = correlation_vector2[best_index];
472 // Compensate index with start offset.
473 best_index = best_index + start_index;
474
475 // Calculate energies.
476 int32_t energy1 = WebRtcSpl_DotProductWithScale(
477 &(audio_history[signal_length - correlation_length]),
478 &(audio_history[signal_length - correlation_length]),
479 correlation_length, correlation_scale);
480 int32_t energy2 = WebRtcSpl_DotProductWithScale(
481 &(audio_history[signal_length - correlation_length - best_index]),
482 &(audio_history[signal_length - correlation_length - best_index]),
483 correlation_length, correlation_scale);
484
485 // Calculate the correlation coefficient between the two portions of the
486 // signal.
487 int32_t corr_coefficient;
488 if ((energy1 > 0) && (energy2 > 0)) {
489 int energy1_scale = std::max(16 - WebRtcSpl_NormW32(energy1), 0);
490 int energy2_scale = std::max(16 - WebRtcSpl_NormW32(energy2), 0);
491 // Make sure total scaling is even (to simplify scale factor after sqrt).
492 if ((energy1_scale + energy2_scale) & 1) {
493 // If sum is odd, add 1 to make it even.
494 energy1_scale += 1;
495 }
496 int32_t scaled_energy1 = energy1 >> energy1_scale;
497 int32_t scaled_energy2 = energy2 >> energy2_scale;
498 int16_t sqrt_energy_product = static_cast<int16_t>(
499 WebRtcSpl_SqrtFloor(scaled_energy1 * scaled_energy2));
500 // Calculate max_correlation / sqrt(energy1 * energy2) in Q14.
501 int cc_shift = 14 - (energy1_scale + energy2_scale) / 2;
502 max_correlation = WEBRTC_SPL_SHIFT_W32(max_correlation, cc_shift);
503 corr_coefficient = WebRtcSpl_DivW32W16(max_correlation,
504 sqrt_energy_product);
505 // Cap at 1.0 in Q14.
506 corr_coefficient = std::min(16384, corr_coefficient);
507 } else {
508 corr_coefficient = 0;
509 }
510
511 // Extract the two vectors expand_vector0 and expand_vector1 from
512 // |audio_history|.
513 size_t expansion_length = max_lag_ + overlap_length_;
514 const int16_t* vector1 = &(audio_history[signal_length - expansion_length]);
515 const int16_t* vector2 = vector1 - distortion_lag;
516 // Normalize the second vector to the same energy as the first.
517 energy1 = WebRtcSpl_DotProductWithScale(vector1, vector1, expansion_length,
518 correlation_scale);
519 energy2 = WebRtcSpl_DotProductWithScale(vector2, vector2, expansion_length,
520 correlation_scale);
521 // Confirm that amplitude ratio sqrt(energy1 / energy2) is within 0.5 - 2.0,
522 // i.e., energy1 / energy2 is within 0.25 - 4.
523 int16_t amplitude_ratio;
524 if ((energy1 / 4 < energy2) && (energy1 > energy2 / 4)) {
525 // Energy constraint fulfilled. Use both vectors and scale them
526 // accordingly.
527 int32_t scaled_energy2 = std::max(16 - WebRtcSpl_NormW32(energy2), 0);
528 int32_t scaled_energy1 = scaled_energy2 - 13;
529 // Calculate scaled_energy1 / scaled_energy2 in Q13.
530 int32_t energy_ratio = WebRtcSpl_DivW32W16(
531 WEBRTC_SPL_SHIFT_W32(energy1, -scaled_energy1),
532 static_cast<int16_t>(energy2 >> scaled_energy2));
533 // Calculate sqrt ratio in Q13 (sqrt of en1/en2 in Q26).
534 amplitude_ratio =
535 static_cast<int16_t>(WebRtcSpl_SqrtFloor(energy_ratio << 13));
536 // Copy the two vectors and give them the same energy.
537 parameters.expand_vector0.Clear();
538 parameters.expand_vector0.PushBack(vector1, expansion_length);
539 parameters.expand_vector1.Clear();
540 if (parameters.expand_vector1.Size() < expansion_length) {
541 parameters.expand_vector1.Extend(
542 expansion_length - parameters.expand_vector1.Size());
543 }
544 WebRtcSpl_AffineTransformVector(¶meters.expand_vector1[0],
545 const_cast<int16_t*>(vector2),
546 amplitude_ratio,
547 4096,
548 13,
549 expansion_length);
550 } else {
551 // Energy change constraint not fulfilled. Only use last vector.
552 parameters.expand_vector0.Clear();
553 parameters.expand_vector0.PushBack(vector1, expansion_length);
554 // Copy from expand_vector0 to expand_vector1.
555 parameters.expand_vector0.CopyTo(¶meters.expand_vector1);
556 // Set the energy_ratio since it is used by muting slope.
557 if ((energy1 / 4 < energy2) || (energy2 == 0)) {
558 amplitude_ratio = 4096; // 0.5 in Q13.
559 } else {
560 amplitude_ratio = 16384; // 2.0 in Q13.
561 }
562 }
563
564 // Set the 3 lag values.
565 if (distortion_lag == correlation_lag) {
566 expand_lags_[0] = distortion_lag;
567 expand_lags_[1] = distortion_lag;
568 expand_lags_[2] = distortion_lag;
569 } else {
570 // |distortion_lag| and |correlation_lag| are not equal; use different
571 // combinations of the two.
572 // First lag is |distortion_lag| only.
573 expand_lags_[0] = distortion_lag;
574 // Second lag is the average of the two.
575 expand_lags_[1] = (distortion_lag + correlation_lag) / 2;
576 // Third lag is the average again, but rounding towards |correlation_lag|.
577 if (distortion_lag > correlation_lag) {
578 expand_lags_[2] = (distortion_lag + correlation_lag - 1) / 2;
579 } else {
580 expand_lags_[2] = (distortion_lag + correlation_lag + 1) / 2;
581 }
582 }
583
584 // Calculate the LPC and the gain of the filters.
585 // Calculate scale value needed for auto-correlation.
586 correlation_scale = WebRtcSpl_MaxAbsValueW16(
587 &(audio_history[signal_length - fs_mult_lpc_analysis_len]),
588 fs_mult_lpc_analysis_len);
589
590 correlation_scale = std::min(16 - WebRtcSpl_NormW32(correlation_scale), 0);
591 correlation_scale = std::max(correlation_scale * 2 + 7, 0);
592
593 // Calculate kUnvoicedLpcOrder + 1 lags of the auto-correlation function.
594 size_t temp_index = signal_length - fs_mult_lpc_analysis_len -
595 kUnvoicedLpcOrder;
596 // Copy signal to temporary vector to be able to pad with leading zeros.
597 int16_t* temp_signal = new int16_t[fs_mult_lpc_analysis_len
598 + kUnvoicedLpcOrder];
599 memset(temp_signal, 0,
600 sizeof(int16_t) * (fs_mult_lpc_analysis_len + kUnvoicedLpcOrder));
601 memcpy(&temp_signal[kUnvoicedLpcOrder],
602 &audio_history[temp_index + kUnvoicedLpcOrder],
603 sizeof(int16_t) * fs_mult_lpc_analysis_len);
604 WebRtcSpl_CrossCorrelation(auto_correlation,
605 &temp_signal[kUnvoicedLpcOrder],
606 &temp_signal[kUnvoicedLpcOrder],
607 fs_mult_lpc_analysis_len, kUnvoicedLpcOrder + 1,
608 correlation_scale, -1);
609 delete [] temp_signal;
610
611 // Verify that variance is positive.
612 if (auto_correlation[0] > 0) {
613 // Estimate AR filter parameters using Levinson-Durbin algorithm;
614 // kUnvoicedLpcOrder + 1 filter coefficients.
615 int16_t stability = WebRtcSpl_LevinsonDurbin(auto_correlation,
616 parameters.ar_filter,
617 reflection_coeff,
618 kUnvoicedLpcOrder);
619
620 // Keep filter parameters only if filter is stable.
621 if (stability != 1) {
622 // Set first coefficient to 4096 (1.0 in Q12).
623 parameters.ar_filter[0] = 4096;
624 // Set remaining |kUnvoicedLpcOrder| coefficients to zero.
625 WebRtcSpl_MemSetW16(parameters.ar_filter + 1, 0, kUnvoicedLpcOrder);
626 }
627 }
628
629 if (channel_ix == 0) {
630 // Extract a noise segment.
631 size_t noise_length;
632 if (distortion_lag < 40) {
633 noise_length = 2 * distortion_lag + 30;
634 } else {
635 noise_length = distortion_lag + 30;
636 }
637 if (noise_length <= RandomVector::kRandomTableSize) {
638 memcpy(random_vector, RandomVector::kRandomTable,
639 sizeof(int16_t) * noise_length);
640 } else {
641 // Only applies to SWB where length could be larger than
642 // |kRandomTableSize|.
643 memcpy(random_vector, RandomVector::kRandomTable,
644 sizeof(int16_t) * RandomVector::kRandomTableSize);
645 assert(noise_length <= kMaxSampleRate / 8000 * 120 + 30);
646 random_vector_->IncreaseSeedIncrement(2);
647 random_vector_->Generate(
648 noise_length - RandomVector::kRandomTableSize,
649 &random_vector[RandomVector::kRandomTableSize]);
650 }
651 }
652
653 // Set up state vector and calculate scale factor for unvoiced filtering.
654 memcpy(parameters.ar_filter_state,
655 &(audio_history[signal_length - kUnvoicedLpcOrder]),
656 sizeof(int16_t) * kUnvoicedLpcOrder);
657 memcpy(unvoiced_vector - kUnvoicedLpcOrder,
658 &(audio_history[signal_length - 128 - kUnvoicedLpcOrder]),
659 sizeof(int16_t) * kUnvoicedLpcOrder);
660 WebRtcSpl_FilterMAFastQ12(&audio_history[signal_length - 128],
661 unvoiced_vector,
662 parameters.ar_filter,
663 kUnvoicedLpcOrder + 1,
664 128);
665 int16_t unvoiced_prescale;
666 if (WebRtcSpl_MaxAbsValueW16(unvoiced_vector, 128) > 4000) {
667 unvoiced_prescale = 4;
668 } else {
669 unvoiced_prescale = 0;
670 }
671 int32_t unvoiced_energy = WebRtcSpl_DotProductWithScale(unvoiced_vector,
672 unvoiced_vector,
673 128,
674 unvoiced_prescale);
675
676 // Normalize |unvoiced_energy| to 28 or 29 bits to preserve sqrt() accuracy.
677 int16_t unvoiced_scale = WebRtcSpl_NormW32(unvoiced_energy) - 3;
678 // Make sure we do an odd number of shifts since we already have 7 shifts
679 // from dividing with 128 earlier. This will make the total scale factor
680 // even, which is suitable for the sqrt.
681 unvoiced_scale += ((unvoiced_scale & 0x1) ^ 0x1);
682 unvoiced_energy = WEBRTC_SPL_SHIFT_W32(unvoiced_energy, unvoiced_scale);
683 int16_t unvoiced_gain =
684 static_cast<int16_t>(WebRtcSpl_SqrtFloor(unvoiced_energy));
685 parameters.ar_gain_scale = 13
686 + (unvoiced_scale + 7 - unvoiced_prescale) / 2;
687 parameters.ar_gain = unvoiced_gain;
688
689 // Calculate voice_mix_factor from corr_coefficient.
690 // Let x = corr_coefficient. Then, we compute:
691 // if (x > 0.48)
692 // voice_mix_factor = (-5179 + 19931x - 16422x^2 + 5776x^3) / 4096;
693 // else
694 // voice_mix_factor = 0;
695 if (corr_coefficient > 7875) {
696 int16_t x1, x2, x3;
697 // |corr_coefficient| is in Q14.
698 x1 = static_cast<int16_t>(corr_coefficient);
699 x2 = (x1 * x1) >> 14; // Shift 14 to keep result in Q14.
700 x3 = (x1 * x2) >> 14;
701 static const int kCoefficients[4] = { -5179, 19931, -16422, 5776 };
702 int32_t temp_sum = kCoefficients[0] << 14;
703 temp_sum += kCoefficients[1] * x1;
704 temp_sum += kCoefficients[2] * x2;
705 temp_sum += kCoefficients[3] * x3;
706 parameters.voice_mix_factor =
707 static_cast<int16_t>(std::min(temp_sum / 4096, 16384));
708 parameters.voice_mix_factor = std::max(parameters.voice_mix_factor,
709 static_cast<int16_t>(0));
710 } else {
711 parameters.voice_mix_factor = 0;
712 }
713
714 // Calculate muting slope. Reuse value from earlier scaling of
715 // |expand_vector0| and |expand_vector1|.
716 int16_t slope = amplitude_ratio;
717 if (slope > 12288) {
718 // slope > 1.5.
719 // Calculate (1 - (1 / slope)) / distortion_lag =
720 // (slope - 1) / (distortion_lag * slope).
721 // |slope| is in Q13, so 1 corresponds to 8192. Shift up to Q25 before
722 // the division.
723 // Shift the denominator from Q13 to Q5 before the division. The result of
724 // the division will then be in Q20.
725 int temp_ratio = WebRtcSpl_DivW32W16(
726 (slope - 8192) << 12,
727 static_cast<int16_t>((distortion_lag * slope) >> 8));
728 if (slope > 14746) {
729 // slope > 1.8.
730 // Divide by 2, with proper rounding.
731 parameters.mute_slope = (temp_ratio + 1) / 2;
732 } else {
733 // Divide by 8, with proper rounding.
734 parameters.mute_slope = (temp_ratio + 4) / 8;
735 }
736 parameters.onset = true;
737 } else {
738 // Calculate (1 - slope) / distortion_lag.
739 // Shift |slope| by 7 to Q20 before the division. The result is in Q20.
740 parameters.mute_slope = WebRtcSpl_DivW32W16(
741 (8192 - slope) << 7, static_cast<int16_t>(distortion_lag));
742 if (parameters.voice_mix_factor <= 13107) {
743 // Make sure the mute factor decreases from 1.0 to 0.9 in no more than
744 // 6.25 ms.
745 // mute_slope >= 0.005 / fs_mult in Q20.
746 parameters.mute_slope = std::max(5243 / fs_mult, parameters.mute_slope);
747 } else if (slope > 8028) {
748 parameters.mute_slope = 0;
749 }
750 parameters.onset = false;
751 }
752 }
753 }
754
ChannelParameters()755 Expand::ChannelParameters::ChannelParameters()
756 : mute_factor(16384),
757 ar_gain(0),
758 ar_gain_scale(0),
759 voice_mix_factor(0),
760 current_voice_mix_factor(0),
761 onset(false),
762 mute_slope(0) {
763 memset(ar_filter, 0, sizeof(ar_filter));
764 memset(ar_filter_state, 0, sizeof(ar_filter_state));
765 }
766
Correlation(const int16_t * input,size_t input_length,int16_t * output,int * output_scale) const767 void Expand::Correlation(const int16_t* input,
768 size_t input_length,
769 int16_t* output,
770 int* output_scale) const {
771 // Set parameters depending on sample rate.
772 const int16_t* filter_coefficients;
773 size_t num_coefficients;
774 int16_t downsampling_factor;
775 if (fs_hz_ == 8000) {
776 num_coefficients = 3;
777 downsampling_factor = 2;
778 filter_coefficients = DspHelper::kDownsample8kHzTbl;
779 } else if (fs_hz_ == 16000) {
780 num_coefficients = 5;
781 downsampling_factor = 4;
782 filter_coefficients = DspHelper::kDownsample16kHzTbl;
783 } else if (fs_hz_ == 32000) {
784 num_coefficients = 7;
785 downsampling_factor = 8;
786 filter_coefficients = DspHelper::kDownsample32kHzTbl;
787 } else { // fs_hz_ == 48000.
788 num_coefficients = 7;
789 downsampling_factor = 12;
790 filter_coefficients = DspHelper::kDownsample48kHzTbl;
791 }
792
793 // Correlate from lag 10 to lag 60 in downsampled domain.
794 // (Corresponds to 20-120 for narrow-band, 40-240 for wide-band, and so on.)
795 static const size_t kCorrelationStartLag = 10;
796 static const size_t kNumCorrelationLags = 54;
797 static const size_t kCorrelationLength = 60;
798 // Downsample to 4 kHz sample rate.
799 static const size_t kDownsampledLength = kCorrelationStartLag
800 + kNumCorrelationLags + kCorrelationLength;
801 int16_t downsampled_input[kDownsampledLength];
802 static const size_t kFilterDelay = 0;
803 WebRtcSpl_DownsampleFast(
804 input + input_length - kDownsampledLength * downsampling_factor,
805 kDownsampledLength * downsampling_factor, downsampled_input,
806 kDownsampledLength, filter_coefficients, num_coefficients,
807 downsampling_factor, kFilterDelay);
808
809 // Normalize |downsampled_input| to using all 16 bits.
810 int16_t max_value = WebRtcSpl_MaxAbsValueW16(downsampled_input,
811 kDownsampledLength);
812 int16_t norm_shift = 16 - WebRtcSpl_NormW32(max_value);
813 WebRtcSpl_VectorBitShiftW16(downsampled_input, kDownsampledLength,
814 downsampled_input, norm_shift);
815
816 int32_t correlation[kNumCorrelationLags];
817 static const int kCorrelationShift = 6;
818 WebRtcSpl_CrossCorrelation(
819 correlation,
820 &downsampled_input[kDownsampledLength - kCorrelationLength],
821 &downsampled_input[kDownsampledLength - kCorrelationLength
822 - kCorrelationStartLag],
823 kCorrelationLength, kNumCorrelationLags, kCorrelationShift, -1);
824
825 // Normalize and move data from 32-bit to 16-bit vector.
826 int32_t max_correlation = WebRtcSpl_MaxAbsValueW32(correlation,
827 kNumCorrelationLags);
828 int16_t norm_shift2 = static_cast<int16_t>(
829 std::max(18 - WebRtcSpl_NormW32(max_correlation), 0));
830 WebRtcSpl_VectorBitShiftW32ToW16(output, kNumCorrelationLags, correlation,
831 norm_shift2);
832 // Total scale factor (right shifts) of correlation value.
833 *output_scale = 2 * norm_shift + kCorrelationShift + norm_shift2;
834 }
835
UpdateLagIndex()836 void Expand::UpdateLagIndex() {
837 current_lag_index_ = current_lag_index_ + lag_index_direction_;
838 // Change direction if needed.
839 if (current_lag_index_ <= 0) {
840 lag_index_direction_ = 1;
841 }
842 if (current_lag_index_ >= kNumLags - 1) {
843 lag_index_direction_ = -1;
844 }
845 }
846
Create(BackgroundNoise * background_noise,SyncBuffer * sync_buffer,RandomVector * random_vector,StatisticsCalculator * statistics,int fs,size_t num_channels) const847 Expand* ExpandFactory::Create(BackgroundNoise* background_noise,
848 SyncBuffer* sync_buffer,
849 RandomVector* random_vector,
850 StatisticsCalculator* statistics,
851 int fs,
852 size_t num_channels) const {
853 return new Expand(background_noise, sync_buffer, random_vector, statistics,
854 fs, num_channels);
855 }
856
857 // 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)858 void Expand::GenerateBackgroundNoise(int16_t* random_vector,
859 size_t channel,
860 int mute_slope,
861 bool too_many_expands,
862 size_t num_noise_samples,
863 int16_t* buffer) {
864 static const size_t kNoiseLpcOrder = BackgroundNoise::kMaxLpcOrder;
865 int16_t scaled_random_vector[kMaxSampleRate / 8000 * 125];
866 assert(num_noise_samples <= (kMaxSampleRate / 8000 * 125));
867 int16_t* noise_samples = &buffer[kNoiseLpcOrder];
868 if (background_noise_->initialized()) {
869 // Use background noise parameters.
870 memcpy(noise_samples - kNoiseLpcOrder,
871 background_noise_->FilterState(channel),
872 sizeof(int16_t) * kNoiseLpcOrder);
873
874 int dc_offset = 0;
875 if (background_noise_->ScaleShift(channel) > 1) {
876 dc_offset = 1 << (background_noise_->ScaleShift(channel) - 1);
877 }
878
879 // Scale random vector to correct energy level.
880 WebRtcSpl_AffineTransformVector(
881 scaled_random_vector, random_vector,
882 background_noise_->Scale(channel), dc_offset,
883 background_noise_->ScaleShift(channel),
884 num_noise_samples);
885
886 WebRtcSpl_FilterARFastQ12(scaled_random_vector, noise_samples,
887 background_noise_->Filter(channel),
888 kNoiseLpcOrder + 1,
889 num_noise_samples);
890
891 background_noise_->SetFilterState(
892 channel,
893 &(noise_samples[num_noise_samples - kNoiseLpcOrder]),
894 kNoiseLpcOrder);
895
896 // Unmute the background noise.
897 int16_t bgn_mute_factor = background_noise_->MuteFactor(channel);
898 NetEq::BackgroundNoiseMode bgn_mode = background_noise_->mode();
899 if (bgn_mode == NetEq::kBgnFade && too_many_expands &&
900 bgn_mute_factor > 0) {
901 // Fade BGN to zero.
902 // Calculate muting slope, approximately -2^18 / fs_hz.
903 int mute_slope;
904 if (fs_hz_ == 8000) {
905 mute_slope = -32;
906 } else if (fs_hz_ == 16000) {
907 mute_slope = -16;
908 } else if (fs_hz_ == 32000) {
909 mute_slope = -8;
910 } else {
911 mute_slope = -5;
912 }
913 // Use UnmuteSignal function with negative slope.
914 // |bgn_mute_factor| is in Q14. |mute_slope| is in Q20.
915 DspHelper::UnmuteSignal(noise_samples,
916 num_noise_samples,
917 &bgn_mute_factor,
918 mute_slope,
919 noise_samples);
920 } else if (bgn_mute_factor < 16384) {
921 // If mode is kBgnOn, or if kBgnFade has started fading,
922 // use regular |mute_slope|.
923 if (!stop_muting_ && bgn_mode != NetEq::kBgnOff &&
924 !(bgn_mode == NetEq::kBgnFade && too_many_expands)) {
925 DspHelper::UnmuteSignal(noise_samples,
926 static_cast<int>(num_noise_samples),
927 &bgn_mute_factor,
928 mute_slope,
929 noise_samples);
930 } else {
931 // kBgnOn and stop muting, or
932 // kBgnOff (mute factor is always 0), or
933 // kBgnFade has reached 0.
934 WebRtcSpl_AffineTransformVector(noise_samples, noise_samples,
935 bgn_mute_factor, 8192, 14,
936 num_noise_samples);
937 }
938 }
939 // Update mute_factor in BackgroundNoise class.
940 background_noise_->SetMuteFactor(channel, bgn_mute_factor);
941 } else {
942 // BGN parameters have not been initialized; use zero noise.
943 memset(noise_samples, 0, sizeof(int16_t) * num_noise_samples);
944 }
945 }
946
GenerateRandomVector(int16_t seed_increment,size_t length,int16_t * random_vector)947 void Expand::GenerateRandomVector(int16_t seed_increment,
948 size_t length,
949 int16_t* random_vector) {
950 // TODO(turajs): According to hlundin The loop should not be needed. Should be
951 // just as good to generate all of the vector in one call.
952 size_t samples_generated = 0;
953 const size_t kMaxRandSamples = RandomVector::kRandomTableSize;
954 while (samples_generated < length) {
955 size_t rand_length = std::min(length - samples_generated, kMaxRandSamples);
956 random_vector_->IncreaseSeedIncrement(seed_increment);
957 random_vector_->Generate(rand_length, &random_vector[samples_generated]);
958 samples_generated += rand_length;
959 }
960 }
961
962 } // namespace webrtc
963