AEC3: Improved the accuracy of the adaptive filter

This CL adds a functionality that jump-starts the
AEC3 shadow filter whenever it performs consistently
worse than the main filter.
The jump-start is done such that the shadow filter
is re-initialized using the main filter coefficients.

The effects of this is a significantly more accurate
main linear filter which leads to less echo leakage
and better transparency

Bug: webrtc:9565, chromium:867873
Change-Id: Ie0b23cd536adc7ce96fc3ed2a7db112aec7437f1
Reviewed-on: https://webrtc-review.googlesource.com/90413
Reviewed-by: Sam Zackrisson <saza@webrtc.org>
Commit-Queue: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#24117}

api/audio/echo_canceller3_config.h

@@ -46,7 +46,7 @@ struct EchoCanceller3Config {
       float noise_gate;
     };
 
-    MainConfiguration main = {13, 0.0005f, 0.01f, 0.001f, 20075344.f};
+    MainConfiguration main = {13, 0.00005f, 0.01f, 0.1f, 20075344.f};
     ShadowConfiguration shadow = {13, 0.7f, 20075344.f};
 
     MainConfiguration main_initial = {12, 0.005f, 0.5f, 0.001f, 20075344.f};

modules/audio_processing/aec3/BUILD.gn

@@ -96,6 +96,7 @@ rtc_static_library("aec3") {
     "stationarity_estimator.h",
     "subtractor.cc",
     "subtractor.h",
+    "subtractor_output.cc",
     "subtractor_output.h",
     "subtractor_output_analyzer.cc",
     "subtractor_output_analyzer.h",

modules/audio_processing/aec3/adaptive_fir_filter.cc

@@ -632,4 +632,14 @@ void AdaptiveFirFilter::ScaleFilter(float factor) {
   }
 }
 
+// Set the filter coefficients.
+void AdaptiveFirFilter::SetFilter(const std::vector<FftData>& H) {
+  RTC_DCHECK_EQ(H_.size(), H.size());
+  const size_t num_partitions = std::min(H_.size(), H.size());
+  for (size_t k = 0; k < num_partitions; ++k) {
+    std::copy(H[k].re.begin(), H[k].re.end(), H_[k].re.begin());
+    std::copy(H[k].im.begin(), H[k].im.end(), H_[k].im.begin());
+  }
+}
+
 }  // namespace webrtc

modules/audio_processing/aec3/adaptive_fir_filter.h

@@ -147,6 +147,12 @@ class AdaptiveFirFilter {
   // Scale the filter impulse response and spectrum by a factor.
   void ScaleFilter(float factor);
 
+  // Set the filter coefficients.
+  void SetFilter(const std::vector<FftData>& H);
+
+  // Gets the filter coefficients.
+  const std::vector<FftData>& GetFilter() const { return H_; }
+
  private:
   // Constrain the filter partitions in a cyclic manner.
   void Constrain();

modules/audio_processing/aec3/aec_state.cc

@@ -123,7 +123,7 @@ void AecState::Update(
     const SubtractorOutput& subtractor_output,
     rtc::ArrayView<const float> y) {
   // Analyze the filter output.
-  subtractor_output_analyzer_.Update(y, subtractor_output);
+  subtractor_output_analyzer_.Update(subtractor_output);
 
   const bool converged_filter = subtractor_output_analyzer_.ConvergedFilter();
   const bool diverged_filter = subtractor_output_analyzer_.DivergedFilter();

modules/audio_processing/aec3/aec_state_unittest.cc

@@ -52,6 +52,7 @@ TEST(AecState, NormalUsage) {
       GetTimeDomainLength(config.filter.main.length_blocks), 0.f);
 
   // Verify that linear AEC usability is false when the filter is diverged.
+  output.UpdatePowers(y);
   state.Update(delay_estimate, diverged_filter_frequency_response,
                impulse_response, *render_delay_buffer->GetRenderBuffer(),
                E2_main, Y2, output, y);
@@ -61,6 +62,7 @@ TEST(AecState, NormalUsage) {
   std::fill(x[0].begin(), x[0].end(), 101.f);
   for (int k = 0; k < 3000; ++k) {
     render_delay_buffer->Insert(x);
+    output.UpdatePowers(y);
     state.Update(delay_estimate, converged_filter_frequency_response,
                  impulse_response, *render_delay_buffer->GetRenderBuffer(),
                  E2_main, Y2, output, y);
@@ -69,6 +71,7 @@ TEST(AecState, NormalUsage) {
 
   // Verify that linear AEC usability becomes false after an echo path change is
   // reported
+  output.UpdatePowers(y);
   state.HandleEchoPathChange(EchoPathVariability(
       false, EchoPathVariability::DelayAdjustment::kBufferReadjustment, false));
   state.Update(delay_estimate, converged_filter_frequency_response,
@@ -79,6 +82,7 @@ TEST(AecState, NormalUsage) {
   // Verify that the active render detection works as intended.
   std::fill(x[0].begin(), x[0].end(), 101.f);
   render_delay_buffer->Insert(x);
+  output.UpdatePowers(y);
   state.HandleEchoPathChange(EchoPathVariability(
       true, EchoPathVariability::DelayAdjustment::kNewDetectedDelay, false));
   state.Update(delay_estimate, converged_filter_frequency_response,
@@ -88,6 +92,7 @@ TEST(AecState, NormalUsage) {
 
   for (int k = 0; k < 1000; ++k) {
     render_delay_buffer->Insert(x);
+    output.UpdatePowers(y);
     state.Update(delay_estimate, converged_filter_frequency_response,
                  impulse_response, *render_delay_buffer->GetRenderBuffer(),
                  E2_main, Y2, output, y);
@@ -111,6 +116,7 @@ TEST(AecState, NormalUsage) {
 
   Y2.fill(10.f * 10000.f * 10000.f);
   for (size_t k = 0; k < 1000; ++k) {
+    output.UpdatePowers(y);
     state.Update(delay_estimate, converged_filter_frequency_response,
                  impulse_response, *render_delay_buffer->GetRenderBuffer(),
                  E2_main, Y2, output, y);
@@ -128,6 +134,7 @@ TEST(AecState, NormalUsage) {
   E2_main.fill(1.f * 10000.f * 10000.f);
   Y2.fill(10.f * E2_main[0]);
   for (size_t k = 0; k < 1000; ++k) {
+    output.UpdatePowers(y);
     state.Update(delay_estimate, converged_filter_frequency_response,
                  impulse_response, *render_delay_buffer->GetRenderBuffer(),
                  E2_main, Y2, output, y);
@@ -149,6 +156,7 @@ TEST(AecState, NormalUsage) {
   E2_main.fill(1.f * 10000.f * 10000.f);
   Y2.fill(5.f * E2_main[0]);
   for (size_t k = 0; k < 1000; ++k) {
+    output.UpdatePowers(y);
     state.Update(delay_estimate, converged_filter_frequency_response,
                  impulse_response, *render_delay_buffer->GetRenderBuffer(),
                  E2_main, Y2, output, y);
@@ -203,6 +211,7 @@ TEST(AecState, ConvergedFilterDelay) {
     impulse_response[k * kBlockSize + 1] = 1.f;
 
     state.HandleEchoPathChange(echo_path_variability);
+    output.UpdatePowers(y);
     state.Update(delay_estimate, frequency_response, impulse_response,
                  *render_delay_buffer->GetRenderBuffer(), E2_main, Y2, output,
                  y);

modules/audio_processing/aec3/echo_canceller3.cc

@@ -50,6 +50,10 @@ bool EnableSlowFilterAdaptation() {
   return !field_trial::IsEnabled("WebRTC-Aec3SlowFilterAdaptationKillSwitch");
 }
 
+bool EnableShadowFilterJumpstart() {
+  return !field_trial::IsEnabled("WebRTC-Aec3ShadowFilterJumpstartKillSwitch");
+}
+
 // Method for adjusting config parameter dependencies..
 EchoCanceller3Config AdjustConfig(const EchoCanceller3Config& config) {
   EchoCanceller3Config adjusted_cfg = config;
@@ -103,12 +107,25 @@ EchoCanceller3Config AdjustConfig(const EchoCanceller3Config& config) {
   }
 
   if (!EnableSlowFilterAdaptation()) {
-    adjusted_cfg.filter.main.leakage_converged = 0.005f;
-    adjusted_cfg.filter.main.leakage_diverged = 0.1f;
+    if (!EnableShadowFilterJumpstart()) {
+      adjusted_cfg.filter.main.leakage_converged = 0.005f;
+      adjusted_cfg.filter.main.leakage_diverged = 0.1f;
+    }
     adjusted_cfg.filter.main_initial.leakage_converged = 0.05f;
     adjusted_cfg.filter.main_initial.leakage_diverged = 5.f;
   }
 
+  if (!EnableShadowFilterJumpstart()) {
+    if (EnableSlowFilterAdaptation()) {
+      adjusted_cfg.filter.main.leakage_converged = 0.0005f;
+      adjusted_cfg.filter.main.leakage_diverged = 0.01f;
+    } else {
+      adjusted_cfg.filter.main.leakage_converged = 0.005f;
+      adjusted_cfg.filter.main.leakage_diverged = 0.1f;
+    }
+    adjusted_cfg.filter.main.error_floor = 0.001f;
+  }
+
   return adjusted_cfg;
 }
 

modules/audio_processing/aec3/subtractor.cc

@@ -36,6 +36,10 @@ bool EnableMisadjustmentEstimator() {
   return !field_trial::IsEnabled("WebRTC-Aec3MisadjustmentEstimatorKillSwitch");
 }
 
+bool EnableShadowFilterJumpstart() {
+  return !field_trial::IsEnabled("WebRTC-Aec3ShadowFilterJumpstartKillSwitch");
+}
+
 void PredictionError(const Aec3Fft& fft,
                      const FftData& S,
                      rtc::ArrayView<const float> y,
@@ -95,6 +99,7 @@ Subtractor::Subtractor(const EchoCanceller3Config& config,
       adaptation_during_saturation_(EnableAdaptationDuringSaturation()),
       enable_misadjustment_estimator_(EnableMisadjustmentEstimator()),
       enable_agc_gain_change_response_(EnableAgcGainChangeResponse()),
+      enable_shadow_filter_jumpstart_(EnableShadowFilterJumpstart()),
       main_filter_(config_.filter.main.length_blocks,
                    config_.filter.main_initial.length_blocks,
                    config.filter.config_change_duration_blocks,
@@ -180,10 +185,13 @@ void Subtractor::Process(const RenderBuffer& render_buffer,
   PredictionError(fft_, S, y, &e_shadow, nullptr, adaptation_during_saturation_,
                   &shadow_saturation);
 
+  // Compute the signal powers in the subtractor output.
+  output->UpdatePowers(y);
+
   // Adjust the filter if needed.
   bool main_filter_adjusted = false;
   if (enable_misadjustment_estimator_) {
-    filter_misadjustment_estimator_.Update(e_main, y);
+    filter_misadjustment_estimator_.Update(*output);
     if (filter_misadjustment_estimator_.IsAdjustmentNeeded()) {
       float scale = filter_misadjustment_estimator_.GetMisadjustment();
       main_filter_.ScaleFilter(scale);
@@ -216,13 +224,22 @@ void Subtractor::Process(const RenderBuffer& render_buffer,
   data_dumper_->DumpRaw("aec3_subtractor_G_main", G.im);
 
   // Update the shadow filter.
-  if (shadow_filter_.SizePartitions() != main_filter_.SizePartitions()) {
-    render_buffer.SpectralSum(shadow_filter_.SizePartitions(), &X2);
+  poor_shadow_filter_counter_ =
+      output->e2_main < output->e2_shadow ? poor_shadow_filter_counter_ + 1 : 0;
+  if (poor_shadow_filter_counter_ < 10 || !enable_shadow_filter_jumpstart_) {
+    if (shadow_filter_.SizePartitions() != main_filter_.SizePartitions()) {
+      render_buffer.SpectralSum(shadow_filter_.SizePartitions(), &X2);
+    }
+    G_shadow_.Compute(X2, render_signal_analyzer, E_shadow,
+                      shadow_filter_.SizePartitions(),
+                      aec_state.SaturatedCapture() || shadow_saturation, &G);
+    shadow_filter_.Adapt(render_buffer, G);
+  } else {
+    G.re.fill(0.f);
+    G.im.fill(0.f);
+    poor_shadow_filter_counter_ = 0;
+    shadow_filter_.SetFilter(main_filter_.GetFilter());
   }
-  G_shadow_.Compute(X2, render_signal_analyzer, E_shadow,
-                    shadow_filter_.SizePartitions(),
-                    aec_state.SaturatedCapture() || shadow_saturation, &G);
-  shadow_filter_.Adapt(render_buffer, G);
 
   data_dumper_->DumpRaw("aec3_subtractor_G_shadow", G.re);
   data_dumper_->DumpRaw("aec3_subtractor_G_shadow", G.im);
@@ -241,19 +258,15 @@ void Subtractor::Process(const RenderBuffer& render_buffer,
 }
 
 void Subtractor::FilterMisadjustmentEstimator::Update(
-    rtc::ArrayView<const float> e,
-    rtc::ArrayView<const float> y) {
-  const auto sum_of_squares = [](float a, float b) { return a + b * b; };
-  const float y2 = std::accumulate(y.begin(), y.end(), 0.f, sum_of_squares);
-  const float e2 = std::accumulate(e.begin(), e.end(), 0.f, sum_of_squares);
-
-  e2_acum_ += e2;
-  y2_acum_ += y2;
+    const SubtractorOutput& output) {
+  e2_acum_ += output.e2_main;
+  y2_acum_ += output.y2;
   if (++n_blocks_acum_ == n_blocks_) {
     if (y2_acum_ > n_blocks_ * 200.f * 200.f * kBlockSize) {
       float update = (e2_acum_ / y2_acum_);
       if (e2_acum_ > n_blocks_ * 7500.f * 7500.f * kBlockSize) {
-        overhang_ = 4;  // Duration equal to blockSizeMs * n_blocks_ * 4
+        // Duration equal to blockSizeMs * n_blocks_ * 4.
+        overhang_ = 4;
       } else {
         overhang_ = std::max(overhang_ - 1, 0);
       }

modules/audio_processing/aec3/subtractor.h

@@ -74,7 +74,7 @@ class Subtractor {
     FilterMisadjustmentEstimator() = default;
     ~FilterMisadjustmentEstimator() = default;
     // Update the misadjustment estimator.
-    void Update(rtc::ArrayView<const float> e, rtc::ArrayView<const float> y);
+    void Update(const SubtractorOutput& output);
     // GetMisadjustment() Returns a recommended scale for the filter so the
     // prediction error energy gets closer to the energy that is seen at the
     // microphone input.
@@ -107,11 +107,13 @@ class Subtractor {
   const bool adaptation_during_saturation_;
   const bool enable_misadjustment_estimator_;
   const bool enable_agc_gain_change_response_;
+  const bool enable_shadow_filter_jumpstart_;
   AdaptiveFirFilter main_filter_;
   AdaptiveFirFilter shadow_filter_;
   MainFilterUpdateGain G_main_;
   ShadowFilterUpdateGain G_shadow_;
   FilterMisadjustmentEstimator filter_misadjustment_estimator_;
+  size_t poor_shadow_filter_counter_ = 0;
   RTC_DISALLOW_IMPLICIT_CONSTRUCTORS(Subtractor);
 };
 

modules/audio_processing/aec3/subtractor_output.cc

@@ -0,0 +1,41 @@
+/*
+ *  Copyright (c) 2018 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/aec3/subtractor_output.h"
+
+#include <numeric>
+
+namespace webrtc {
+
+SubtractorOutput::SubtractorOutput() = default;
+SubtractorOutput::~SubtractorOutput() = default;
+
+void SubtractorOutput::Reset() {
+  s_main.fill(0.f);
+  e_main.fill(0.f);
+  e_shadow.fill(0.f);
+  E_main.re.fill(0.f);
+  E_main.im.fill(0.f);
+  E2_main.fill(0.f);
+  E2_shadow.fill(0.f);
+  e2_main = 0.f;
+  e2_shadow = 0.f;
+  y2 = 0.f;
+}
+
+void SubtractorOutput::UpdatePowers(rtc::ArrayView<const float> y) {
+  const auto sum_of_squares = [](float a, float b) { return a + b * b; };
+  y2 = std::accumulate(y.begin(), y.end(), 0.f, sum_of_squares);
+  e2_main = std::accumulate(e_main.begin(), e_main.end(), 0.f, sum_of_squares);
+  e2_shadow =
+      std::accumulate(e_shadow.begin(), e_shadow.end(), 0.f, sum_of_squares);
+}
+
+}  // namespace webrtc

modules/audio_processing/aec3/subtractor_output.h

@@ -13,6 +13,7 @@
 
 #include <array>
 
+#include "api/array_view.h"
 #include "modules/audio_processing/aec3/aec3_common.h"
 #include "modules/audio_processing/aec3/fft_data.h"
 
@@ -20,22 +21,24 @@ namespace webrtc {
 
 // Stores the values being returned from the echo subtractor.
 struct SubtractorOutput {
+  SubtractorOutput();
+  ~SubtractorOutput();
+
   std::array<float, kBlockSize> s_main;
   std::array<float, kBlockSize> e_main;
   std::array<float, kBlockSize> e_shadow;
   FftData E_main;
   std::array<float, kFftLengthBy2Plus1> E2_main;
   std::array<float, kFftLengthBy2Plus1> E2_shadow;
+  float e2_main = 0.f;
+  float e2_shadow = 0.f;
+  float y2 = 0.f;
 
-  void Reset() {
-    s_main.fill(0.f);
-    e_main.fill(0.f);
-    e_shadow.fill(0.f);
-    E_main.re.fill(0.f);
-    E_main.im.fill(0.f);
-    E2_main.fill(0.f);
-    E2_shadow.fill(0.f);
-  }
+  // Reset the struct content.
+  void Reset();
+
+  // Updates the powers of the signals.
+  void UpdatePowers(rtc::ArrayView<const float> y);
 };
 
 }  // namespace webrtc

modules/audio_processing/aec3/subtractor_output_analyzer.cc

@@ -16,17 +16,10 @@
 namespace webrtc {
 
 void SubtractorOutputAnalyzer::Update(
-    rtc::ArrayView<const float> y,
     const SubtractorOutput& subtractor_output) {
-  const auto& e_main = subtractor_output.e_main;
-  const auto& e_shadow = subtractor_output.e_shadow;
-
-  const auto sum_of_squares = [](float a, float b) { return a + b * b; };
-  const float y2 = std::accumulate(y.begin(), y.end(), 0.f, sum_of_squares);
-  const float e2_main =
-      std::accumulate(e_main.begin(), e_main.end(), 0.f, sum_of_squares);
-  const float e2_shadow =
-      std::accumulate(e_shadow.begin(), e_shadow.end(), 0.f, sum_of_squares);
+  const float y2 = subtractor_output.y2;
+  const float e2_main = subtractor_output.e2_main;
+  const float e2_shadow = subtractor_output.e2_shadow;
 
   constexpr float kConvergenceThreshold = 50 * 50 * kBlockSize;
   main_filter_converged_ = e2_main < 0.5f * y2 && y2 > kConvergenceThreshold;

modules/audio_processing/aec3/subtractor_output_analyzer.h

@@ -23,8 +23,7 @@ class SubtractorOutputAnalyzer {
   ~SubtractorOutputAnalyzer() = default;
 
   // Analyses the subtractor output.
-  void Update(rtc::ArrayView<const float> y,
-              const SubtractorOutput& subtractor_output);
+  void Update(const SubtractorOutput& subtractor_output);
 
   bool ConvergedFilter() const {
     return main_filter_converged_ || shadow_filter_converged_;