AEC3: Increasing the transparency during call startup

This CL increases the AEC3 transparency during call
startup and after echo path delay changes in 3 ways:
1. The exit requirements for the initial mode is
made less strict.
2. The requirements for using the linear echo model
are made less strict.
3. The duplicated reverb modelling in the linear mode
removed.


Bug: webrtc:9572,chromium:868329
Change-Id: I79ea0796ed26408e35576bb39eaae4e4848b4f83
Reviewed-on: https://webrtc-review.googlesource.com/90868
Reviewed-by: Sam Zackrisson <saza@webrtc.org>
Commit-Queue: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#24132}

modules/audio_processing/aec3/aec_state.cc

@@ -45,6 +45,14 @@ bool EnableLinearModeWithDivergedFilter() {
       "WebRTC-Aec3LinearModeWithDivergedFilterKillSwitch");
 }
 
+bool EnableEarlyFilterUsage() {
+  return !field_trial::IsEnabled("WebRTC-Aec3EarlyLinearFilterUsageKillSwitch");
+}
+
+bool EnableShortInitialState() {
+  return !field_trial::IsEnabled("WebRTC-Aec3ShortInitialStateKillSwitch");
+}
+
 float ComputeGainRampupIncrease(const EchoCanceller3Config& config) {
   const auto& c = config.echo_removal_control.gain_rampup;
   return powf(1.f / c.first_non_zero_gain, 1.f / c.non_zero_gain_blocks);
@@ -68,6 +76,8 @@ AecState::AecState(const EchoCanceller3Config& config)
       enforce_delay_after_realignment_(EnableEnforcingDelayAfterRealignment()),
       allow_linear_mode_with_diverged_filter_(
           EnableLinearModeWithDivergedFilter()),
+      early_filter_usage_activated_(EnableEarlyFilterUsage()),
+      use_short_initial_state_(EnableShortInitialState()),
       erle_estimator_(config.erle.min, config.erle.max_l, config.erle.max_h),
       max_render_(config_.filter.main.length_blocks, 0.f),
       gain_rampup_increase_(ComputeGainRampupIncrease(config_)),
@@ -190,8 +200,14 @@ void AecState::Update(
     echo_saturation_ = DetectEchoSaturation(x, EchoPathGain());
   }
 
-  bool filter_has_had_time_to_converge =
+  bool filter_has_had_time_to_converge;
+  if (early_filter_usage_activated_) {
+    filter_has_had_time_to_converge =
+        blocks_with_proper_filter_adaptation_ >= 0.8f * kNumBlocksPerSecond;
+  } else {
+    filter_has_had_time_to_converge =
         blocks_with_proper_filter_adaptation_ >= 1.5f * kNumBlocksPerSecond;
+  }
 
   if (!filter_should_have_converged_) {
     filter_should_have_converged_ =
@@ -199,8 +215,13 @@ void AecState::Update(
   }
 
   // Flag whether the initial state is still active.
+  if (use_short_initial_state_) {
+    initial_state_ =
+        blocks_with_proper_filter_adaptation_ < 2.5f * kNumBlocksPerSecond;
+  } else {
     initial_state_ =
         blocks_with_proper_filter_adaptation_ < 5 * kNumBlocksPerSecond;
+  }
 
   // Update counters for the filter divergence and convergence.
   diverged_blocks_ = diverged_filter ? diverged_blocks_ + 1 : 0;

modules/audio_processing/aec3/aec_state.h

@@ -182,6 +182,8 @@ class AecState {
   const bool use_stationary_properties_;
   const bool enforce_delay_after_realignment_;
   const bool allow_linear_mode_with_diverged_filter_;
+  const bool early_filter_usage_activated_;
+  const bool use_short_initial_state_;
   ErlEstimator erl_estimator_;
   ErleEstimator erle_estimator_;
   size_t capture_block_counter_ = 0;

modules/audio_processing/aec3/residual_echo_estimator.cc

@@ -30,6 +30,11 @@ bool OverrideEstimatedEchoPathGain() {
   return !field_trial::IsEnabled("WebRTC-Aec3OverrideEchoPathGainKillSwitch");
 }
 
+bool UseFixedNonLinearReverbModel() {
+  return field_trial::IsEnabled(
+      "WebRTC-Aec3StandardNonlinearReverbModelKillSwitch");
+}
+
 // Computes the indexes that will be used for computing spectral power over
 // the blocks surrounding the delay.
 void GetRenderIndexesToAnalyze(
@@ -74,7 +79,8 @@ void GetRenderIndexesToAnalyze(
 ResidualEchoEstimator::ResidualEchoEstimator(const EchoCanceller3Config& config)
     : config_(config),
       soft_transparent_mode_(EnableSoftTransparentMode()),
-      override_estimated_echo_path_gain_(OverrideEstimatedEchoPathGain()) {
+      override_estimated_echo_path_gain_(OverrideEstimatedEchoPathGain()),
+      use_fixed_nonlinear_reverb_model_(UseFixedNonLinearReverbModel()) {
   if (config_.ep_strength.reverb_based_on_render) {
     echo_reverb_.reset(new ReverbModel());
   } else {
@@ -230,6 +236,7 @@ void ResidualEchoEstimator::NonLinearEstimate(
     return a * echo_path_gain * echo_path_gain;
   });
 
+  if (use_fixed_nonlinear_reverb_model_) {
     for (size_t k = 0; k < R2->size(); ++k) {
       // Update hold counter.
       R2_hold_counter_[k] = R2_old_[k] < (*R2)[k] ? 0 : R2_hold_counter_[k] + 1;
@@ -239,11 +246,12 @@ void ResidualEchoEstimator::NonLinearEstimate(
       (*R2)[k] =
           R2_hold_counter_[k] < config_.echo_model.nonlinear_hold
               ? std::max((*R2)[k], R2_old_[k])
-            : std::min(
-                  (*R2)[k] + R2_old_[k] * config_.echo_model.nonlinear_release,
+              : std::min((*R2)[k] +
+                             R2_old_[k] * config_.echo_model.nonlinear_release,
                          Y2[k]);
     }
   }
+}
 
 void ResidualEchoEstimator::EchoGeneratingPower(
     const VectorBuffer& spectrum_buffer,

modules/audio_processing/aec3/residual_echo_estimator.h

@@ -91,6 +91,7 @@ class ResidualEchoEstimator {
   std::array<int, kFftLengthBy2Plus1> X2_noise_floor_counter_;
   const bool soft_transparent_mode_;
   const bool override_estimated_echo_path_gain_;
+  const bool use_fixed_nonlinear_reverb_model_;
   std::unique_ptr<ReverbModel> echo_reverb_;
   std::unique_ptr<ReverbModelFallback> echo_reverb_fallback;
   RTC_DISALLOW_IMPLICIT_CONSTRUCTORS(ResidualEchoEstimator);