• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  Copyright (c) 2014 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 //
12 //  Implements core class for intelligibility enhancer.
13 //
14 //  Details of the model and algorithm can be found in the original paper:
15 //  http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6882788
16 //
17 
18 #include "webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.h"
19 
20 #include <math.h>
21 #include <stdlib.h>
22 #include <algorithm>
23 #include <numeric>
24 
25 #include "webrtc/base/checks.h"
26 #include "webrtc/common_audio/include/audio_util.h"
27 #include "webrtc/common_audio/window_generator.h"
28 
29 namespace webrtc {
30 
31 namespace {
32 
33 const size_t kErbResolution = 2;
34 const int kWindowSizeMs = 2;
35 const int kChunkSizeMs = 10;  // Size provided by APM.
36 const float kClipFreq = 200.0f;
37 const float kConfigRho = 0.02f;  // Default production and interpretation SNR.
38 const float kKbdAlpha = 1.5f;
39 const float kLambdaBot = -1.0f;      // Extreme values in bisection
40 const float kLambdaTop = -10e-18f;  // search for lamda.
41 
42 }  // namespace
43 
44 using std::complex;
45 using std::max;
46 using std::min;
47 using VarianceType = intelligibility::VarianceArray::StepType;
48 
TransformCallback(IntelligibilityEnhancer * parent,IntelligibilityEnhancer::AudioSource source)49 IntelligibilityEnhancer::TransformCallback::TransformCallback(
50     IntelligibilityEnhancer* parent,
51     IntelligibilityEnhancer::AudioSource source)
52     : parent_(parent), source_(source) {
53 }
54 
ProcessAudioBlock(const complex<float> * const * in_block,size_t in_channels,size_t frames,size_t,complex<float> * const * out_block)55 void IntelligibilityEnhancer::TransformCallback::ProcessAudioBlock(
56     const complex<float>* const* in_block,
57     size_t in_channels,
58     size_t frames,
59     size_t /* out_channels */,
60     complex<float>* const* out_block) {
61   RTC_DCHECK_EQ(parent_->freqs_, frames);
62   for (size_t i = 0; i < in_channels; ++i) {
63     parent_->DispatchAudio(source_, in_block[i], out_block[i]);
64   }
65 }
66 
IntelligibilityEnhancer()67 IntelligibilityEnhancer::IntelligibilityEnhancer()
68     : IntelligibilityEnhancer(IntelligibilityEnhancer::Config()) {
69 }
70 
IntelligibilityEnhancer(const Config & config)71 IntelligibilityEnhancer::IntelligibilityEnhancer(const Config& config)
72     : freqs_(RealFourier::ComplexLength(
73           RealFourier::FftOrder(config.sample_rate_hz * kWindowSizeMs / 1000))),
74       window_size_(static_cast<size_t>(1 << RealFourier::FftOrder(freqs_))),
75       chunk_length_(
76           static_cast<size_t>(config.sample_rate_hz * kChunkSizeMs / 1000)),
77       bank_size_(GetBankSize(config.sample_rate_hz, kErbResolution)),
78       sample_rate_hz_(config.sample_rate_hz),
79       erb_resolution_(kErbResolution),
80       num_capture_channels_(config.num_capture_channels),
81       num_render_channels_(config.num_render_channels),
82       analysis_rate_(config.analysis_rate),
83       active_(true),
84       clear_variance_(freqs_,
85                       config.var_type,
86                       config.var_window_size,
87                       config.var_decay_rate),
88       noise_variance_(freqs_,
89                       config.var_type,
90                       config.var_window_size,
91                       config.var_decay_rate),
92       filtered_clear_var_(new float[bank_size_]),
93       filtered_noise_var_(new float[bank_size_]),
94       filter_bank_(bank_size_),
95       center_freqs_(new float[bank_size_]),
96       rho_(new float[bank_size_]),
97       gains_eq_(new float[bank_size_]),
98       gain_applier_(freqs_, config.gain_change_limit),
99       temp_render_out_buffer_(chunk_length_, num_render_channels_),
100       temp_capture_out_buffer_(chunk_length_, num_capture_channels_),
101       kbd_window_(new float[window_size_]),
102       render_callback_(this, AudioSource::kRenderStream),
103       capture_callback_(this, AudioSource::kCaptureStream),
104       block_count_(0),
105       analysis_step_(0) {
106   RTC_DCHECK_LE(config.rho, 1.0f);
107 
108   CreateErbBank();
109 
110   // Assumes all rho equal.
111   for (size_t i = 0; i < bank_size_; ++i) {
112     rho_[i] = config.rho * config.rho;
113   }
114 
115   float freqs_khz = kClipFreq / 1000.0f;
116   size_t erb_index = static_cast<size_t>(ceilf(
117       11.17f * logf((freqs_khz + 0.312f) / (freqs_khz + 14.6575f)) + 43.0f));
118   start_freq_ = std::max(static_cast<size_t>(1), erb_index * erb_resolution_);
119 
120   WindowGenerator::KaiserBesselDerived(kKbdAlpha, window_size_,
121                                        kbd_window_.get());
122   render_mangler_.reset(new LappedTransform(
123       num_render_channels_, num_render_channels_, chunk_length_,
124       kbd_window_.get(), window_size_, window_size_ / 2, &render_callback_));
125   capture_mangler_.reset(new LappedTransform(
126       num_capture_channels_, num_capture_channels_, chunk_length_,
127       kbd_window_.get(), window_size_, window_size_ / 2, &capture_callback_));
128 }
129 
ProcessRenderAudio(float * const * audio,int sample_rate_hz,size_t num_channels)130 void IntelligibilityEnhancer::ProcessRenderAudio(float* const* audio,
131                                                  int sample_rate_hz,
132                                                  size_t num_channels) {
133   RTC_CHECK_EQ(sample_rate_hz_, sample_rate_hz);
134   RTC_CHECK_EQ(num_render_channels_, num_channels);
135 
136   if (active_) {
137     render_mangler_->ProcessChunk(audio, temp_render_out_buffer_.channels());
138   }
139 
140   if (active_) {
141     for (size_t i = 0; i < num_render_channels_; ++i) {
142       memcpy(audio[i], temp_render_out_buffer_.channels()[i],
143              chunk_length_ * sizeof(**audio));
144     }
145   }
146 }
147 
AnalyzeCaptureAudio(float * const * audio,int sample_rate_hz,size_t num_channels)148 void IntelligibilityEnhancer::AnalyzeCaptureAudio(float* const* audio,
149                                                   int sample_rate_hz,
150                                                   size_t num_channels) {
151   RTC_CHECK_EQ(sample_rate_hz_, sample_rate_hz);
152   RTC_CHECK_EQ(num_capture_channels_, num_channels);
153 
154   capture_mangler_->ProcessChunk(audio, temp_capture_out_buffer_.channels());
155 }
156 
DispatchAudio(IntelligibilityEnhancer::AudioSource source,const complex<float> * in_block,complex<float> * out_block)157 void IntelligibilityEnhancer::DispatchAudio(
158     IntelligibilityEnhancer::AudioSource source,
159     const complex<float>* in_block,
160     complex<float>* out_block) {
161   switch (source) {
162     case kRenderStream:
163       ProcessClearBlock(in_block, out_block);
164       break;
165     case kCaptureStream:
166       ProcessNoiseBlock(in_block, out_block);
167       break;
168   }
169 }
170 
ProcessClearBlock(const complex<float> * in_block,complex<float> * out_block)171 void IntelligibilityEnhancer::ProcessClearBlock(const complex<float>* in_block,
172                                                 complex<float>* out_block) {
173   if (block_count_ < 2) {
174     memset(out_block, 0, freqs_ * sizeof(*out_block));
175     ++block_count_;
176     return;
177   }
178 
179   // TODO(ekm): Use VAD to |Step| and |AnalyzeClearBlock| only if necessary.
180   if (true) {
181     clear_variance_.Step(in_block, false);
182     if (block_count_ % analysis_rate_ == analysis_rate_ - 1) {
183       const float power_target = std::accumulate(
184           clear_variance_.variance(), clear_variance_.variance() + freqs_, 0.f);
185       AnalyzeClearBlock(power_target);
186       ++analysis_step_;
187     }
188     ++block_count_;
189   }
190 
191   if (active_) {
192     gain_applier_.Apply(in_block, out_block);
193   }
194 }
195 
AnalyzeClearBlock(float power_target)196 void IntelligibilityEnhancer::AnalyzeClearBlock(float power_target) {
197   FilterVariance(clear_variance_.variance(), filtered_clear_var_.get());
198   FilterVariance(noise_variance_.variance(), filtered_noise_var_.get());
199 
200   SolveForGainsGivenLambda(kLambdaTop, start_freq_, gains_eq_.get());
201   const float power_top =
202       DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
203   SolveForGainsGivenLambda(kLambdaBot, start_freq_, gains_eq_.get());
204   const float power_bot =
205       DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
206   if (power_target >= power_bot && power_target <= power_top) {
207     SolveForLambda(power_target, power_bot, power_top);
208     UpdateErbGains();
209   }  // Else experiencing variance underflow, so do nothing.
210 }
211 
SolveForLambda(float power_target,float power_bot,float power_top)212 void IntelligibilityEnhancer::SolveForLambda(float power_target,
213                                              float power_bot,
214                                              float power_top) {
215   const float kConvergeThresh = 0.001f;  // TODO(ekmeyerson): Find best values
216   const int kMaxIters = 100;             // for these, based on experiments.
217 
218   const float reciprocal_power_target = 1.f / power_target;
219   float lambda_bot = kLambdaBot;
220   float lambda_top = kLambdaTop;
221   float power_ratio = 2.0f;  // Ratio of achieved power to target power.
222   int iters = 0;
223   while (std::fabs(power_ratio - 1.0f) > kConvergeThresh &&
224          iters <= kMaxIters) {
225     const float lambda = lambda_bot + (lambda_top - lambda_bot) / 2.0f;
226     SolveForGainsGivenLambda(lambda, start_freq_, gains_eq_.get());
227     const float power =
228         DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
229     if (power < power_target) {
230       lambda_bot = lambda;
231     } else {
232       lambda_top = lambda;
233     }
234     power_ratio = std::fabs(power * reciprocal_power_target);
235     ++iters;
236   }
237 }
238 
UpdateErbGains()239 void IntelligibilityEnhancer::UpdateErbGains() {
240   // (ERB gain) = filterbank' * (freq gain)
241   float* gains = gain_applier_.target();
242   for (size_t i = 0; i < freqs_; ++i) {
243     gains[i] = 0.0f;
244     for (size_t j = 0; j < bank_size_; ++j) {
245       gains[i] = fmaf(filter_bank_[j][i], gains_eq_[j], gains[i]);
246     }
247   }
248 }
249 
ProcessNoiseBlock(const complex<float> * in_block,complex<float> *)250 void IntelligibilityEnhancer::ProcessNoiseBlock(const complex<float>* in_block,
251                                                 complex<float>* /*out_block*/) {
252   noise_variance_.Step(in_block);
253 }
254 
GetBankSize(int sample_rate,size_t erb_resolution)255 size_t IntelligibilityEnhancer::GetBankSize(int sample_rate,
256                                             size_t erb_resolution) {
257   float freq_limit = sample_rate / 2000.0f;
258   size_t erb_scale = static_cast<size_t>(ceilf(
259       11.17f * logf((freq_limit + 0.312f) / (freq_limit + 14.6575f)) + 43.0f));
260   return erb_scale * erb_resolution;
261 }
262 
CreateErbBank()263 void IntelligibilityEnhancer::CreateErbBank() {
264   size_t lf = 1, rf = 4;
265 
266   for (size_t i = 0; i < bank_size_; ++i) {
267     float abs_temp = fabsf((i + 1.0f) / static_cast<float>(erb_resolution_));
268     center_freqs_[i] = 676170.4f / (47.06538f - expf(0.08950404f * abs_temp));
269     center_freqs_[i] -= 14678.49f;
270   }
271   float last_center_freq = center_freqs_[bank_size_ - 1];
272   for (size_t i = 0; i < bank_size_; ++i) {
273     center_freqs_[i] *= 0.5f * sample_rate_hz_ / last_center_freq;
274   }
275 
276   for (size_t i = 0; i < bank_size_; ++i) {
277     filter_bank_[i].resize(freqs_);
278   }
279 
280   for (size_t i = 1; i <= bank_size_; ++i) {
281     size_t lll, ll, rr, rrr;
282     static const size_t kOne = 1;  // Avoids repeated static_cast<>s below.
283     lll = static_cast<size_t>(round(
284         center_freqs_[max(kOne, i - lf) - 1] * freqs_ /
285             (0.5f * sample_rate_hz_)));
286     ll = static_cast<size_t>(round(
287         center_freqs_[max(kOne, i) - 1] * freqs_ / (0.5f * sample_rate_hz_)));
288     lll = min(freqs_, max(lll, kOne)) - 1;
289     ll = min(freqs_, max(ll, kOne)) - 1;
290 
291     rrr = static_cast<size_t>(round(
292         center_freqs_[min(bank_size_, i + rf) - 1] * freqs_ /
293             (0.5f * sample_rate_hz_)));
294     rr = static_cast<size_t>(round(
295         center_freqs_[min(bank_size_, i + 1) - 1] * freqs_ /
296             (0.5f * sample_rate_hz_)));
297     rrr = min(freqs_, max(rrr, kOne)) - 1;
298     rr = min(freqs_, max(rr, kOne)) - 1;
299 
300     float step, element;
301 
302     step = 1.0f / (ll - lll);
303     element = 0.0f;
304     for (size_t j = lll; j <= ll; ++j) {
305       filter_bank_[i - 1][j] = element;
306       element += step;
307     }
308     step = 1.0f / (rrr - rr);
309     element = 1.0f;
310     for (size_t j = rr; j <= rrr; ++j) {
311       filter_bank_[i - 1][j] = element;
312       element -= step;
313     }
314     for (size_t j = ll; j <= rr; ++j) {
315       filter_bank_[i - 1][j] = 1.0f;
316     }
317   }
318 
319   float sum;
320   for (size_t i = 0; i < freqs_; ++i) {
321     sum = 0.0f;
322     for (size_t j = 0; j < bank_size_; ++j) {
323       sum += filter_bank_[j][i];
324     }
325     for (size_t j = 0; j < bank_size_; ++j) {
326       filter_bank_[j][i] /= sum;
327     }
328   }
329 }
330 
SolveForGainsGivenLambda(float lambda,size_t start_freq,float * sols)331 void IntelligibilityEnhancer::SolveForGainsGivenLambda(float lambda,
332                                                        size_t start_freq,
333                                                        float* sols) {
334   bool quadratic = (kConfigRho < 1.0f);
335   const float* var_x0 = filtered_clear_var_.get();
336   const float* var_n0 = filtered_noise_var_.get();
337 
338   for (size_t n = 0; n < start_freq; ++n) {
339     sols[n] = 1.0f;
340   }
341 
342   // Analytic solution for optimal gains. See paper for derivation.
343   for (size_t n = start_freq - 1; n < bank_size_; ++n) {
344     float alpha0, beta0, gamma0;
345     gamma0 = 0.5f * rho_[n] * var_x0[n] * var_n0[n] +
346              lambda * var_x0[n] * var_n0[n] * var_n0[n];
347     beta0 = lambda * var_x0[n] * (2 - rho_[n]) * var_x0[n] * var_n0[n];
348     if (quadratic) {
349       alpha0 = lambda * var_x0[n] * (1 - rho_[n]) * var_x0[n] * var_x0[n];
350       sols[n] =
351           (-beta0 - sqrtf(beta0 * beta0 - 4 * alpha0 * gamma0)) / (2 * alpha0);
352     } else {
353       sols[n] = -gamma0 / beta0;
354     }
355     sols[n] = fmax(0, sols[n]);
356   }
357 }
358 
FilterVariance(const float * var,float * result)359 void IntelligibilityEnhancer::FilterVariance(const float* var, float* result) {
360   RTC_DCHECK_GT(freqs_, 0u);
361   for (size_t i = 0; i < bank_size_; ++i) {
362     result[i] = DotProduct(&filter_bank_[i][0], var, freqs_);
363   }
364 }
365 
DotProduct(const float * a,const float * b,size_t length)366 float IntelligibilityEnhancer::DotProduct(const float* a,
367                                           const float* b,
368                                           size_t length) {
369   float ret = 0.0f;
370 
371   for (size_t i = 0; i < length; ++i) {
372     ret = fmaf(a[i], b[i], ret);
373   }
374   return ret;
375 }
376 
active() const377 bool IntelligibilityEnhancer::active() const {
378   return active_;
379 }
380 
381 }  // namespace webrtc
382