1 /*
2  *  Copyright (c) 2016 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_processing/residual_echo_detector.h"
12 
13 #include <algorithm>
14 #include <numeric>
15 
16 #include "modules/audio_processing/audio_buffer.h"
17 #include "modules/audio_processing/logging/apm_data_dumper.h"
18 #include "rtc_base/atomicops.h"
19 #include "rtc_base/logging.h"
20 #include "system_wrappers/include/metrics.h"
21 
22 namespace {
23 
Power(rtc::ArrayView<const float> input)24 float Power(rtc::ArrayView<const float> input) {
25   if (input.empty()) {
26     return 0.f;
27   }
28   return std::inner_product(input.begin(), input.end(), input.begin(), 0.f) /
29          input.size();
30 }
31 
32 constexpr size_t kLookbackFrames = 650;
33 // TODO(ivoc): Verify the size of this buffer.
34 constexpr size_t kRenderBufferSize = 30;
35 constexpr float kAlpha = 0.001f;
36 // 10 seconds of data, updated every 10 ms.
37 constexpr size_t kAggregationBufferSize = 10 * 100;
38 
39 }  // namespace
40 
41 namespace webrtc {
42 
43 int ResidualEchoDetector::instance_count_ = 0;
44 
ResidualEchoDetector()45 ResidualEchoDetector::ResidualEchoDetector()
46     : data_dumper_(
47           new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
48       render_buffer_(kRenderBufferSize),
49       render_power_(kLookbackFrames),
50       render_power_mean_(kLookbackFrames),
51       render_power_std_dev_(kLookbackFrames),
52       covariances_(kLookbackFrames),
53       recent_likelihood_max_(kAggregationBufferSize) {}
54 
55 ResidualEchoDetector::~ResidualEchoDetector() = default;
56 
AnalyzeRenderAudio(rtc::ArrayView<const float> render_audio)57 void ResidualEchoDetector::AnalyzeRenderAudio(
58     rtc::ArrayView<const float> render_audio) {
59   // Dump debug data assuming 48 kHz sample rate (if this assumption is not
60   // valid the dumped audio will need to be converted offline accordingly).
61   data_dumper_->DumpWav("ed_render", render_audio.size(), render_audio.data(),
62                         48000, 1);
63 
64   if (render_buffer_.Size() == 0) {
65     frames_since_zero_buffer_size_ = 0;
66   } else if (frames_since_zero_buffer_size_ >= kRenderBufferSize) {
67     // This can happen in a few cases: at the start of a call, due to a glitch
68     // or due to clock drift. The excess capture value will be ignored.
69     // TODO(ivoc): Include how often this happens in APM stats.
70     render_buffer_.Pop();
71     frames_since_zero_buffer_size_ = 0;
72   }
73   ++frames_since_zero_buffer_size_;
74   float power = Power(render_audio);
75   render_buffer_.Push(power);
76 }
77 
AnalyzeCaptureAudio(rtc::ArrayView<const float> capture_audio)78 void ResidualEchoDetector::AnalyzeCaptureAudio(
79     rtc::ArrayView<const float> capture_audio) {
80   // Dump debug data assuming 48 kHz sample rate (if this assumption is not
81   // valid the dumped audio will need to be converted offline accordingly).
82   data_dumper_->DumpWav("ed_capture", capture_audio.size(),
83                         capture_audio.data(), 48000, 1);
84 
85   if (first_process_call_) {
86     // On the first process call (so the start of a call), we must flush the
87     // render buffer, otherwise the render data will be delayed.
88     render_buffer_.Clear();
89     first_process_call_ = false;
90   }
91 
92   // Get the next render value.
93   const rtc::Optional<float> buffered_render_power = render_buffer_.Pop();
94   if (!buffered_render_power) {
95     // This can happen in a few cases: at the start of a call, due to a glitch
96     // or due to clock drift. The excess capture value will be ignored.
97     // TODO(ivoc): Include how often this happens in APM stats.
98     return;
99   }
100   // Update the render statistics, and store the statistics in circular buffers.
101   render_statistics_.Update(*buffered_render_power);
102   RTC_DCHECK_LT(next_insertion_index_, kLookbackFrames);
103   render_power_[next_insertion_index_] = *buffered_render_power;
104   render_power_mean_[next_insertion_index_] = render_statistics_.mean();
105   render_power_std_dev_[next_insertion_index_] =
106       render_statistics_.std_deviation();
107 
108   // Get the next capture value, update capture statistics and add the relevant
109   // values to the buffers.
110   const float capture_power = Power(capture_audio);
111   capture_statistics_.Update(capture_power);
112   const float capture_mean = capture_statistics_.mean();
113   const float capture_std_deviation = capture_statistics_.std_deviation();
114 
115   // Update the covariance values and determine the new echo likelihood.
116   echo_likelihood_ = 0.f;
117   size_t read_index = next_insertion_index_;
118 
119   int best_delay = -1;
120   for (size_t delay = 0; delay < covariances_.size(); ++delay) {
121     RTC_DCHECK_LT(read_index, render_power_.size());
122     covariances_[delay].Update(capture_power, capture_mean,
123                                capture_std_deviation, render_power_[read_index],
124                                render_power_mean_[read_index],
125                                render_power_std_dev_[read_index]);
126     read_index = read_index > 0 ? read_index - 1 : kLookbackFrames - 1;
127 
128     if (covariances_[delay].normalized_cross_correlation() > echo_likelihood_) {
129       echo_likelihood_ = covariances_[delay].normalized_cross_correlation();
130       best_delay = static_cast<int>(delay);
131     }
132   }
133   // This is a temporary log message to help find the underlying cause for echo
134   // likelihoods > 1.0.
135   // TODO(ivoc): Remove once the issue is resolved.
136   if (echo_likelihood_ > 1.1f) {
137     // Make sure we don't spam the log.
138     if (log_counter_ < 5 && best_delay != -1) {
139       size_t read_index = kLookbackFrames + next_insertion_index_ - best_delay;
140       if (read_index >= kLookbackFrames) {
141         read_index -= kLookbackFrames;
142       }
143       RTC_DCHECK_LT(read_index, render_power_.size());
144       RTC_LOG_F(LS_ERROR) << "Echo detector internal state: {"
145                           << "Echo likelihood: " << echo_likelihood_
146                           << ", Best Delay: " << best_delay << ", Covariance: "
147                           << covariances_[best_delay].covariance()
148                           << ", Last capture power: " << capture_power
149                           << ", Capture mean: " << capture_mean
150                           << ", Capture_standard deviation: "
151                           << capture_std_deviation << ", Last render power: "
152                           << render_power_[read_index]
153                           << ", Render mean: " << render_power_mean_[read_index]
154                           << ", Render standard deviation: "
155                           << render_power_std_dev_[read_index]
156                           << ", Reliability: " << reliability_ << "}";
157       log_counter_++;
158     }
159   }
160   RTC_DCHECK_LT(echo_likelihood_, 1.1f);
161 
162   reliability_ = (1.0f - kAlpha) * reliability_ + kAlpha * 1.0f;
163   echo_likelihood_ *= reliability_;
164   // This is a temporary fix to prevent echo likelihood values > 1.0.
165   // TODO(ivoc): Find the root cause of this issue and fix it.
166   echo_likelihood_ = std::min(echo_likelihood_, 1.0f);
167   int echo_percentage = static_cast<int>(echo_likelihood_ * 100);
168   RTC_HISTOGRAM_COUNTS("WebRTC.Audio.ResidualEchoDetector.EchoLikelihood",
169                        echo_percentage, 0, 100, 100 /* number of bins */);
170 
171   // Update the buffer of recent likelihood values.
172   recent_likelihood_max_.Update(echo_likelihood_);
173 
174   // Update the next insertion index.
175   next_insertion_index_ = next_insertion_index_ < (kLookbackFrames - 1)
176                               ? next_insertion_index_ + 1
177                               : 0;
178 }
179 
Initialize()180 void ResidualEchoDetector::Initialize() {
181   render_buffer_.Clear();
182   std::fill(render_power_.begin(), render_power_.end(), 0.f);
183   std::fill(render_power_mean_.begin(), render_power_mean_.end(), 0.f);
184   std::fill(render_power_std_dev_.begin(), render_power_std_dev_.end(), 0.f);
185   render_statistics_.Clear();
186   capture_statistics_.Clear();
187   recent_likelihood_max_.Clear();
188   for (auto& cov : covariances_) {
189     cov.Clear();
190   }
191   echo_likelihood_ = 0.f;
192   next_insertion_index_ = 0;
193   reliability_ = 0.f;
194 }
195 
PackRenderAudioBuffer(AudioBuffer * audio,std::vector<float> * packed_buffer)196 void ResidualEchoDetector::PackRenderAudioBuffer(
197     AudioBuffer* audio,
198     std::vector<float>* packed_buffer) {
199   packed_buffer->clear();
200   packed_buffer->insert(packed_buffer->end(), audio->channels_f()[0],
201                         audio->channels_f()[0] + audio->num_frames());
202 }
203 
204 }  // namespace webrtc
205