mirror of
https://github.com/mollyim/webrtc.git
synced 2025-05-14 06:10:40 +01:00

Important: This change does not in any way affect echo cancellation or standardized stats. The user audio experience is unchanged. Only non-standard stats are affected. Echo return loss metrics are unchanged. Residual echo likelihood {recent max} will no longer be computed by default. Important: The echo detector is no longer enabled by default. API change, PSA: https://groups.google.com/g/discuss-webrtc/c/mJV5cDysBDI/m/7PTPBjVHCgAJ This CL removes the default usage of the residual echo detector in APM. It can now only be used via injection and the helper function webrtc::CreateEchoDetector. See how the function audio_processing_unittest.cc:CreateApm() changed, for an example. The echo detector implementation is marked poisonous, to avoid accidental dependencies. Some cleanup is done: - EchoDetector::PackRenderAudioBuffer is declared in one target but is defined in another target. It is not necessary to keep in the API. It is made an implementation detail, and the echo detector input is documented in the API. - The internal state of APM is large and difficult to track. Submodule pointers that are set permanently on construction are now appropriately marked const. Tested: - existing + new unit tests - audioproc_f is bitexact on a large number of aecdumps Bug: webrtc:11539 Change-Id: I00cc2ee112fedb06451a533409311605220064d0 Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/239652 Reviewed-by: Ivo Creusen <ivoc@webrtc.org> Reviewed-by: Per Kjellander <perkj@webrtc.org> Reviewed-by: Mirko Bonadei <mbonadei@webrtc.org> Commit-Queue: Sam Zackrisson <saza@webrtc.org> Cr-Commit-Position: refs/heads/main@{#35550}
207 lines
8.3 KiB
C++
207 lines
8.3 KiB
C++
/*
|
|
* Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
|
|
*
|
|
* Use of this source code is governed by a BSD-style license
|
|
* that can be found in the LICENSE file in the root of the source
|
|
* tree. An additional intellectual property rights grant can be found
|
|
* in the file PATENTS. All contributing project authors may
|
|
* be found in the AUTHORS file in the root of the source tree.
|
|
*/
|
|
|
|
#include "modules/audio_processing/residual_echo_detector.h"
|
|
|
|
#include <algorithm>
|
|
#include <numeric>
|
|
|
|
#include "absl/types/optional.h"
|
|
#include "modules/audio_processing/logging/apm_data_dumper.h"
|
|
#include "rtc_base/atomic_ops.h"
|
|
#include "rtc_base/checks.h"
|
|
#include "rtc_base/logging.h"
|
|
#include "system_wrappers/include/metrics.h"
|
|
|
|
namespace {
|
|
|
|
float Power(rtc::ArrayView<const float> input) {
|
|
if (input.empty()) {
|
|
return 0.f;
|
|
}
|
|
return std::inner_product(input.begin(), input.end(), input.begin(), 0.f) /
|
|
input.size();
|
|
}
|
|
|
|
constexpr size_t kLookbackFrames = 650;
|
|
// TODO(ivoc): Verify the size of this buffer.
|
|
constexpr size_t kRenderBufferSize = 30;
|
|
constexpr float kAlpha = 0.001f;
|
|
// 10 seconds of data, updated every 10 ms.
|
|
constexpr size_t kAggregationBufferSize = 10 * 100;
|
|
|
|
} // namespace
|
|
|
|
namespace webrtc {
|
|
|
|
int ResidualEchoDetector::instance_count_ = 0;
|
|
|
|
ResidualEchoDetector::ResidualEchoDetector()
|
|
: data_dumper_(
|
|
new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
|
|
render_buffer_(kRenderBufferSize),
|
|
render_power_(kLookbackFrames),
|
|
render_power_mean_(kLookbackFrames),
|
|
render_power_std_dev_(kLookbackFrames),
|
|
covariances_(kLookbackFrames),
|
|
recent_likelihood_max_(kAggregationBufferSize) {}
|
|
|
|
ResidualEchoDetector::~ResidualEchoDetector() = default;
|
|
|
|
void ResidualEchoDetector::AnalyzeRenderAudio(
|
|
rtc::ArrayView<const float> render_audio) {
|
|
// Dump debug data assuming 48 kHz sample rate (if this assumption is not
|
|
// valid the dumped audio will need to be converted offline accordingly).
|
|
data_dumper_->DumpWav("ed_render", render_audio.size(), render_audio.data(),
|
|
48000, 1);
|
|
|
|
if (render_buffer_.Size() == 0) {
|
|
frames_since_zero_buffer_size_ = 0;
|
|
} else if (frames_since_zero_buffer_size_ >= kRenderBufferSize) {
|
|
// This can happen in a few cases: at the start of a call, due to a glitch
|
|
// or due to clock drift. The excess capture value will be ignored.
|
|
// TODO(ivoc): Include how often this happens in APM stats.
|
|
render_buffer_.Pop();
|
|
frames_since_zero_buffer_size_ = 0;
|
|
}
|
|
++frames_since_zero_buffer_size_;
|
|
float power = Power(render_audio);
|
|
render_buffer_.Push(power);
|
|
}
|
|
|
|
void ResidualEchoDetector::AnalyzeCaptureAudio(
|
|
rtc::ArrayView<const float> capture_audio) {
|
|
// Dump debug data assuming 48 kHz sample rate (if this assumption is not
|
|
// valid the dumped audio will need to be converted offline accordingly).
|
|
data_dumper_->DumpWav("ed_capture", capture_audio.size(),
|
|
capture_audio.data(), 48000, 1);
|
|
|
|
if (first_process_call_) {
|
|
// On the first process call (so the start of a call), we must flush the
|
|
// render buffer, otherwise the render data will be delayed.
|
|
render_buffer_.Clear();
|
|
first_process_call_ = false;
|
|
}
|
|
|
|
// Get the next render value.
|
|
const absl::optional<float> buffered_render_power = render_buffer_.Pop();
|
|
if (!buffered_render_power) {
|
|
// This can happen in a few cases: at the start of a call, due to a glitch
|
|
// or due to clock drift. The excess capture value will be ignored.
|
|
// TODO(ivoc): Include how often this happens in APM stats.
|
|
return;
|
|
}
|
|
// Update the render statistics, and store the statistics in circular buffers.
|
|
render_statistics_.Update(*buffered_render_power);
|
|
RTC_DCHECK_LT(next_insertion_index_, kLookbackFrames);
|
|
render_power_[next_insertion_index_] = *buffered_render_power;
|
|
render_power_mean_[next_insertion_index_] = render_statistics_.mean();
|
|
render_power_std_dev_[next_insertion_index_] =
|
|
render_statistics_.std_deviation();
|
|
|
|
// Get the next capture value, update capture statistics and add the relevant
|
|
// values to the buffers.
|
|
const float capture_power = Power(capture_audio);
|
|
capture_statistics_.Update(capture_power);
|
|
const float capture_mean = capture_statistics_.mean();
|
|
const float capture_std_deviation = capture_statistics_.std_deviation();
|
|
|
|
// Update the covariance values and determine the new echo likelihood.
|
|
echo_likelihood_ = 0.f;
|
|
size_t read_index = next_insertion_index_;
|
|
|
|
int best_delay = -1;
|
|
for (size_t delay = 0; delay < covariances_.size(); ++delay) {
|
|
RTC_DCHECK_LT(read_index, render_power_.size());
|
|
covariances_[delay].Update(capture_power, capture_mean,
|
|
capture_std_deviation, render_power_[read_index],
|
|
render_power_mean_[read_index],
|
|
render_power_std_dev_[read_index]);
|
|
read_index = read_index > 0 ? read_index - 1 : kLookbackFrames - 1;
|
|
|
|
if (covariances_[delay].normalized_cross_correlation() > echo_likelihood_) {
|
|
echo_likelihood_ = covariances_[delay].normalized_cross_correlation();
|
|
best_delay = static_cast<int>(delay);
|
|
}
|
|
}
|
|
// This is a temporary log message to help find the underlying cause for echo
|
|
// likelihoods > 1.0.
|
|
// TODO(ivoc): Remove once the issue is resolved.
|
|
if (echo_likelihood_ > 1.1f) {
|
|
// Make sure we don't spam the log.
|
|
if (log_counter_ < 5 && best_delay != -1) {
|
|
size_t read_index = kLookbackFrames + next_insertion_index_ - best_delay;
|
|
if (read_index >= kLookbackFrames) {
|
|
read_index -= kLookbackFrames;
|
|
}
|
|
RTC_DCHECK_LT(read_index, render_power_.size());
|
|
RTC_LOG_F(LS_ERROR) << "Echo detector internal state: {"
|
|
"Echo likelihood: "
|
|
<< echo_likelihood_ << ", Best Delay: " << best_delay
|
|
<< ", Covariance: "
|
|
<< covariances_[best_delay].covariance()
|
|
<< ", Last capture power: " << capture_power
|
|
<< ", Capture mean: " << capture_mean
|
|
<< ", Capture_standard deviation: "
|
|
<< capture_std_deviation << ", Last render power: "
|
|
<< render_power_[read_index]
|
|
<< ", Render mean: " << render_power_mean_[read_index]
|
|
<< ", Render standard deviation: "
|
|
<< render_power_std_dev_[read_index]
|
|
<< ", Reliability: " << reliability_ << "}";
|
|
log_counter_++;
|
|
}
|
|
}
|
|
RTC_DCHECK_LT(echo_likelihood_, 1.1f);
|
|
|
|
reliability_ = (1.0f - kAlpha) * reliability_ + kAlpha * 1.0f;
|
|
echo_likelihood_ *= reliability_;
|
|
// This is a temporary fix to prevent echo likelihood values > 1.0.
|
|
// TODO(ivoc): Find the root cause of this issue and fix it.
|
|
echo_likelihood_ = std::min(echo_likelihood_, 1.0f);
|
|
int echo_percentage = static_cast<int>(echo_likelihood_ * 100);
|
|
RTC_HISTOGRAM_COUNTS("WebRTC.Audio.ResidualEchoDetector.EchoLikelihood",
|
|
echo_percentage, 0, 100, 100 /* number of bins */);
|
|
|
|
// Update the buffer of recent likelihood values.
|
|
recent_likelihood_max_.Update(echo_likelihood_);
|
|
|
|
// Update the next insertion index.
|
|
next_insertion_index_ = next_insertion_index_ < (kLookbackFrames - 1)
|
|
? next_insertion_index_ + 1
|
|
: 0;
|
|
}
|
|
|
|
void ResidualEchoDetector::Initialize(int /*capture_sample_rate_hz*/,
|
|
int /*num_capture_channels*/,
|
|
int /*render_sample_rate_hz*/,
|
|
int /*num_render_channels*/) {
|
|
render_buffer_.Clear();
|
|
std::fill(render_power_.begin(), render_power_.end(), 0.f);
|
|
std::fill(render_power_mean_.begin(), render_power_mean_.end(), 0.f);
|
|
std::fill(render_power_std_dev_.begin(), render_power_std_dev_.end(), 0.f);
|
|
render_statistics_.Clear();
|
|
capture_statistics_.Clear();
|
|
recent_likelihood_max_.Clear();
|
|
for (auto& cov : covariances_) {
|
|
cov.Clear();
|
|
}
|
|
echo_likelihood_ = 0.f;
|
|
next_insertion_index_ = 0;
|
|
reliability_ = 0.f;
|
|
}
|
|
|
|
EchoDetector::Metrics ResidualEchoDetector::GetMetrics() const {
|
|
EchoDetector::Metrics metrics;
|
|
metrics.echo_likelihood = echo_likelihood_;
|
|
metrics.echo_likelihood_recent_max = recent_likelihood_max_.max();
|
|
return metrics;
|
|
}
|
|
} // namespace webrtc
|