AEC3: Kill kill-switches

"Perfection is achieved, not when there is nothing more to add,
but when there is nothing left to take away."

This CL removes the following kill-switches from AEC3
- WebRTC-Aec3DownSamplingFactor8KillSwitch
- WebRTC-Aec3NewSuppressionKillSwitch
- WebRTC-Aec3ShadowFilterJumpstartKillSwitch
- WebRTC-Aec3SlowFilterAdaptationKillSwitch
- WebRTC-Aec3SuppressorNearendAveragingKillSwitch

It also removes code paths and configuration parameters that are no
longer in use. The list of kill-switches in the audio processing
fuzzer test is updated.

The change has been tested for bit-exactness.

Bug: webrtc:8671
Change-Id: Ie0af86a14baf853548bf9c00b2b9b3bbc32c1aaa
Reviewed-on: https://webrtc-review.googlesource.com/c/105324
Reviewed-by: Sam Zackrisson <saza@webrtc.org>
Reviewed-by: Per Åhgren <peah@webrtc.org>
Commit-Queue: Gustaf Ullberg <gustaf@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#25120}

api/audio/echo_canceller3_config.cc

@@ -45,9 +45,6 @@ EchoCanceller3Config::Delay::Delay() = default;
 EchoCanceller3Config::Delay::Delay(const EchoCanceller3Config::Delay& e) =
     default;
 
-EchoCanceller3Config::Mask::Mask() = default;
-EchoCanceller3Config::Mask::Mask(const EchoCanceller3Config::Mask& m) = default;
-
 EchoCanceller3Config::EchoModel::EchoModel() = default;
 EchoCanceller3Config::EchoModel::EchoModel(
     const EchoCanceller3Config::EchoModel& e) = default;
@@ -158,22 +155,6 @@ bool EchoCanceller3Config::Validate(EchoCanceller3Config* config) {
   res = res && Limit(&c->ep_strength.hf, 0.f, 1000000.f);
   res = res && Limit(&c->ep_strength.default_len, 0.f, 1.f);
 
-  res = res && Limit(&c->gain_mask.m0, 0.f, 1.f);
-  res = res && Limit(&c->gain_mask.m1, 0.f, 1.f);
-  res = res && Limit(&c->gain_mask.m2, 0.f, 1.f);
-  res = res && Limit(&c->gain_mask.m3, 0.f, 1.f);
-  res = res && Limit(&c->gain_mask.m5, 0.f, 1.f);
-  res = res && Limit(&c->gain_mask.m6, 0.f, 1.f);
-  res = res && Limit(&c->gain_mask.m7, 0.f, 1.f);
-  res = res && Limit(&c->gain_mask.m8, 0.f, 1.f);
-  res = res && Limit(&c->gain_mask.m9, 0.f, 1.f);
-
-  res = res && Limit(&c->gain_mask.gain_curve_offset, 0.f, 1000.f);
-  res = res && Limit(&c->gain_mask.gain_curve_slope, 0.f, 1000.f);
-  res = res && Limit(&c->gain_mask.temporal_masking_lf, 0.f, 1000.f);
-  res = res && Limit(&c->gain_mask.temporal_masking_hf, 0.f, 1000.f);
-  res = res && Limit(&c->gain_mask.temporal_masking_lf_bands, 0, 65);
-
   res = res &&
         Limit(&c->echo_audibility.low_render_limit, 0.f, 32768.f * 32768.f);
   res = res &&

api/audio/echo_canceller3_config.h

@@ -92,26 +92,6 @@ struct RTC_EXPORT EchoCanceller3Config {
     bool bounded_erl = false;
   } ep_strength;
 
-  struct Mask {
-    Mask();
-    Mask(const Mask& m);
-    float m0 = 0.1f;
-    float m1 = 0.01f;
-    float m2 = 0.0001f;
-    float m3 = 0.01f;
-    float m5 = 0.01f;
-    float m6 = 0.0001f;
-    float m7 = 0.01f;
-    float m8 = 0.0001f;
-    float m9 = 0.1f;
-
-    float gain_curve_offset = 1.45f;
-    float gain_curve_slope = 5.f;
-    float temporal_masking_lf = 0.9f;
-    float temporal_masking_hf = 0.6f;
-    size_t temporal_masking_lf_bands = 3;
-  } gain_mask;
-
   struct EchoAudibility {
     float low_render_limit = 4 * 64.f;
     float normal_render_limit = 64.f;

api/audio/echo_canceller3_config_json.cc

@@ -192,26 +192,6 @@ EchoCanceller3Config Aec3ConfigFromJsonString(absl::string_view json_string) {
     ReadParam(section, "bounded_erl", &cfg.ep_strength.bounded_erl);
   }
 
-  if (rtc::GetValueFromJsonObject(aec3_root, "gain_mask", &section)) {
-    ReadParam(section, "m1", &cfg.gain_mask.m1);
-    ReadParam(section, "m2", &cfg.gain_mask.m2);
-    ReadParam(section, "m3", &cfg.gain_mask.m3);
-    ReadParam(section, "m5", &cfg.gain_mask.m5);
-    ReadParam(section, "m6", &cfg.gain_mask.m6);
-    ReadParam(section, "m7", &cfg.gain_mask.m7);
-    ReadParam(section, "m8", &cfg.gain_mask.m8);
-    ReadParam(section, "m9", &cfg.gain_mask.m9);
-
-    ReadParam(section, "gain_curve_offset", &cfg.gain_mask.gain_curve_offset);
-    ReadParam(section, "gain_curve_slope", &cfg.gain_mask.gain_curve_slope);
-    ReadParam(section, "temporal_masking_lf",
-              &cfg.gain_mask.temporal_masking_lf);
-    ReadParam(section, "temporal_masking_hf",
-              &cfg.gain_mask.temporal_masking_hf);
-    ReadParam(section, "temporal_masking_lf_bands",
-              &cfg.gain_mask.temporal_masking_lf_bands);
-  }
-
   if (rtc::GetValueFromJsonObject(aec3_root, "echo_audibility", &section)) {
     ReadParam(section, "low_render_limit",
               &cfg.echo_audibility.low_render_limit);
@@ -424,26 +404,6 @@ std::string Aec3ConfigToJsonString(const EchoCanceller3Config& config) {
 
   ost << "},";
 
-  ost << "\"gain_mask\": {";
-  ost << "\"m0\": " << config.gain_mask.m0 << ",";
-  ost << "\"m1\": " << config.gain_mask.m1 << ",";
-  ost << "\"m2\": " << config.gain_mask.m2 << ",";
-  ost << "\"m3\": " << config.gain_mask.m3 << ",";
-  ost << "\"m5\": " << config.gain_mask.m5 << ",";
-  ost << "\"m6\": " << config.gain_mask.m6 << ",";
-  ost << "\"m7\": " << config.gain_mask.m7 << ",";
-  ost << "\"m8\": " << config.gain_mask.m8 << ",";
-  ost << "\"m9\": " << config.gain_mask.m9 << ",";
-  ost << "\"gain_curve_offset\": " << config.gain_mask.gain_curve_offset << ",";
-  ost << "\"gain_curve_slope\": " << config.gain_mask.gain_curve_slope << ",";
-  ost << "\"temporal_masking_lf\": " << config.gain_mask.temporal_masking_lf
-      << ",";
-  ost << "\"temporal_masking_hf\": " << config.gain_mask.temporal_masking_hf
-      << ",";
-  ost << "\"temporal_masking_lf_bands\": "
-      << config.gain_mask.temporal_masking_lf_bands;
-  ost << "},";
-
   ost << "\"echo_audibility\": {";
   ost << "\"low_render_limit\": " << config.echo_audibility.low_render_limit
       << ",";

modules/audio_processing/aec3/echo_canceller3.cc

@@ -41,19 +41,6 @@ bool EnableReverbModelling() {
   return !field_trial::IsEnabled("WebRTC-Aec3ReverbModellingKillSwitch");
 }
 
-bool EnableSuppressorNearendAveraging() {
-  return !field_trial::IsEnabled(
-      "WebRTC-Aec3SuppressorNearendAveragingKillSwitch");
-}
-
-bool EnableSlowFilterAdaptation() {
-  return !field_trial::IsEnabled("WebRTC-Aec3SlowFilterAdaptationKillSwitch");
-}
-
-bool EnableShadowFilterJumpstart() {
-  return !field_trial::IsEnabled("WebRTC-Aec3ShadowFilterJumpstartKillSwitch");
-}
-
 bool EnableUnityInitialRampupGain() {
   return field_trial::IsEnabled("WebRTC-Aec3EnableUnityInitialRampupGain");
 }
@@ -101,30 +88,6 @@ EchoCanceller3Config AdjustConfig(const EchoCanceller3Config& config) {
     adjusted_cfg.ep_strength.reverb_based_on_render = false;
   }
 
-  if (!EnableSuppressorNearendAveraging()) {
-    adjusted_cfg.suppressor.nearend_average_blocks = 1;
-  }
-
-  if (!EnableSlowFilterAdaptation()) {
-    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;
-  }
-
   if (!EnableNewFilterParams()) {
     adjusted_cfg.filter.main.leakage_diverged = 0.01f;
     adjusted_cfg.filter.main.error_floor = 0.1f;

modules/audio_processing/aec3/echo_path_delay_estimator.cc

@@ -19,21 +19,12 @@
 #include "system_wrappers/include/field_trial.h"
 
 namespace webrtc {
-namespace {
-size_t GetDownSamplingFactor(const EchoCanceller3Config& config) {
-  // Do not use down sampling factor 8 if kill switch is triggered.
-  return (config.delay.down_sampling_factor == 8 &&
-          field_trial::IsEnabled("WebRTC-Aec3DownSamplingFactor8KillSwitch"))
-             ? 4
-             : config.delay.down_sampling_factor;
-}
-}  // namespace
 
 EchoPathDelayEstimator::EchoPathDelayEstimator(
     ApmDataDumper* data_dumper,
     const EchoCanceller3Config& config)
     : data_dumper_(data_dumper),
-      down_sampling_factor_(GetDownSamplingFactor(config)),
+      down_sampling_factor_(config.delay.down_sampling_factor),
       sub_block_size_(down_sampling_factor_ != 0
                           ? kBlockSize / down_sampling_factor_
                           : kBlockSize),
@@ -45,7 +36,7 @@ EchoPathDelayEstimator::EchoPathDelayEstimator(
           kMatchedFilterWindowSizeSubBlocks,
           config.delay.num_filters,
           kMatchedFilterAlignmentShiftSizeSubBlocks,
-          GetDownSamplingFactor(config) == 8
+          config.delay.down_sampling_factor == 8
               ? config.render_levels.poor_excitation_render_limit_ds8
               : config.render_levels.poor_excitation_render_limit,
           config.delay.delay_estimate_smoothing,

modules/audio_processing/aec3/render_delay_buffer.cc

@@ -35,14 +35,6 @@ bool EnableZeroExternalDelayHeadroom() {
       "WebRTC-Aec3ZeroExternalDelayHeadroomKillSwitch");
 }
 
-size_t GetDownSamplingFactor(const EchoCanceller3Config& config) {
-  // Do not use down sampling factor 8 if kill switch is triggered.
-  return (config.delay.down_sampling_factor == 8 &&
-          field_trial::IsEnabled("WebRTC-Aec3DownSamplingFactor8KillSwitch"))
-             ? 4
-             : config.delay.down_sampling_factor;
-}
-
 class RenderDelayBufferImpl final : public RenderDelayBuffer {
  public:
   RenderDelayBufferImpl(const EchoCanceller3Config& config, size_t num_bands);
@@ -177,7 +169,7 @@ RenderDelayBufferImpl::RenderDelayBufferImpl(const EchoCanceller3Config& config,
           new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
       optimization_(DetectOptimization()),
       config_(config),
-      down_sampling_factor_(GetDownSamplingFactor(config)),
+      down_sampling_factor_(config.delay.down_sampling_factor),
       use_zero_external_delay_headroom_(EnableZeroExternalDelayHeadroom()),
       sub_block_size_(static_cast<int>(down_sampling_factor_ > 0
                                            ? kBlockSize / down_sampling_factor_

modules/audio_processing/aec3/suppression_gain.cc

@@ -25,10 +25,6 @@
 namespace webrtc {
 namespace {
 
-bool EnableNewSuppression() {
-  return !field_trial::IsEnabled("WebRTC-Aec3NewSuppressionKillSwitch");
-}
-
 // Adjust the gains according to the presence of known external filters.
 void AdjustForExternalFilters(std::array<float, kFftLengthBy2Plus1>* gain) {
   // Limit the low frequency gains to avoid the impact of the high-pass filter
@@ -83,57 +79,6 @@ void WeightEchoForAudibility(const EchoCanceller3Config& config,
   weigh(threshold, normalizer, 7, kFftLengthBy2Plus1, echo, weighted_echo);
 }
 
-// Computes the gain to reduce the echo to a non audible level.
-void GainToNoAudibleEchoFallback(
-    const EchoCanceller3Config& config,
-    bool low_noise_render,
-    bool saturated_echo,
-    bool linear_echo_estimate,
-    const std::array<float, kFftLengthBy2Plus1>& nearend,
-    const std::array<float, kFftLengthBy2Plus1>& weighted_echo,
-    const std::array<float, kFftLengthBy2Plus1>& masker,
-    const std::array<float, kFftLengthBy2Plus1>& min_gain,
-    const std::array<float, kFftLengthBy2Plus1>& max_gain,
-    const std::array<float, kFftLengthBy2Plus1>& one_by_weighted_echo,
-    std::array<float, kFftLengthBy2Plus1>* gain) {
-  float nearend_masking_margin = 0.f;
-  if (linear_echo_estimate) {
-    nearend_masking_margin =
-        low_noise_render
-            ? config.gain_mask.m9
-            : (saturated_echo ? config.gain_mask.m2 : config.gain_mask.m3);
-  } else {
-    nearend_masking_margin = config.gain_mask.m7;
-  }
-
-  RTC_DCHECK_LE(0.f, nearend_masking_margin);
-  RTC_DCHECK_GT(1.f, nearend_masking_margin);
-
-  const float masker_margin =
-      linear_echo_estimate ? config.gain_mask.m0 : config.gain_mask.m8;
-
-  for (size_t k = 0; k < gain->size(); ++k) {
-    // TODO(devicentepena): Experiment by removing the reverberation estimation
-    // from the nearend signal before computing the gains.
-    const float unity_gain_masker = std::max(nearend[k], masker[k]);
-    RTC_DCHECK_LE(0.f, nearend_masking_margin * unity_gain_masker);
-    if (weighted_echo[k] <= nearend_masking_margin * unity_gain_masker ||
-        unity_gain_masker <= 0.f) {
-      (*gain)[k] = 1.f;
-    } else {
-      RTC_DCHECK_LT(0.f, unity_gain_masker);
-      (*gain)[k] =
-          std::max(0.f, (1.f - config.gain_mask.gain_curve_slope *
-                                   weighted_echo[k] / unity_gain_masker) *
-                            config.gain_mask.gain_curve_offset);
-      (*gain)[k] = std::max(masker_margin * masker[k] * one_by_weighted_echo[k],
-                            (*gain)[k]);
-    }
-
-    (*gain)[k] = std::min(std::max((*gain)[k], min_gain[k]), max_gain[k]);
-  }
-}
-
 // TODO(peah): Make adaptive to take the actual filter error into account.
 constexpr size_t kUpperAccurateBandPlus1 = 29;
 
@@ -321,30 +266,9 @@ void SuppressionGain::LowerBandGain(
   std::array<float, kFftLengthBy2Plus1> max_gain;
   GetMaxGain(max_gain);
 
-  // Iteratively compute the gain required to attenuate the echo to a non
-  // noticeable level.
-
-  if (enable_new_suppression_) {
     GainToNoAudibleEcho(nearend, weighted_residual_echo, comfort_noise,
                         min_gain, max_gain, gain);
     AdjustForExternalFilters(gain);
-  } else {
-    const bool linear_echo_estimate = aec_state.UsableLinearEstimate();
-    std::array<float, kFftLengthBy2Plus1> masker;
-    std::array<float, kFftLengthBy2Plus1> one_by_weighted_echo;
-    std::transform(weighted_residual_echo.begin(), weighted_residual_echo.end(),
-                   one_by_weighted_echo.begin(),
-                   [](float e) { return e > 0.f ? 1.f / e : 1.f; });
-    gain->fill(0.f);
-    for (int k = 0; k < 2; ++k) {
-      std::copy(comfort_noise.begin(), comfort_noise.end(), masker.begin());
-      GainToNoAudibleEchoFallback(config_, low_noise_render, saturated_echo,
-                                  linear_echo_estimate, nearend,
-                                  weighted_residual_echo, masker, min_gain,
-                                  max_gain, one_by_weighted_echo, gain);
-      AdjustForExternalFilters(gain);
-    }
-  }
 
   // Adjust the gain for frequencies which have not yet converged.
   AdjustNonConvergedFrequencies(gain);
@@ -372,7 +296,6 @@ SuppressionGain::SuppressionGain(const EchoCanceller3Config& config,
       config_(config),
       state_change_duration_blocks_(
           static_cast<int>(config_.filter.config_change_duration_blocks)),
-      enable_new_suppression_(EnableNewSuppression()),
       moving_average_(kFftLengthBy2Plus1,
                       config.suppressor.nearend_average_blocks),
       nearend_params_(config_.suppressor.nearend_tuning),

modules/audio_processing/aec3/suppression_gain.h

@@ -137,7 +137,6 @@ class SuppressionGain {
   LowNoiseRenderDetector low_render_detector_;
   bool initial_state_ = true;
   int initial_state_change_counter_ = 0;
-  const bool enable_new_suppression_;
   aec3::MovingAverage moving_average_;
   const GainParameters nearend_params_;
   const GainParameters normal_params_;

test/fuzzers/audio_processing_configs_fuzzer.cc

@@ -25,33 +25,39 @@ namespace webrtc {
 namespace {
 
 const std::string kFieldTrialNames[] = {
-    "WebRTC-Aec3TransparentModeKillSwitch",
-    "WebRTC-Aec3StationaryRenderImprovementsKillSwitch",
-    "WebRTC-Aec3EnforceDelayAfterRealignmentKillSwitch",
-    "WebRTC-Aec3UseShortDelayEstimatorWindow",
-    "WebRTC-Aec3ReverbBasedOnRenderKillSwitch",
-    "WebRTC-Aec3ReverbModellingKillSwitch",
-    "WebRTC-Aec3FilterAnalyzerPreprocessorKillSwitch",
-    "WebRTC-Aec3TransparencyImprovementsKillSwitch",
-    "WebRTC-Aec3SoftTransparentModeKillSwitch",
-    "WebRTC-Aec3OverrideEchoPathGainKillSwitch",
-    "WebRTC-Aec3ZeroExternalDelayHeadroomKillSwitch",
-    "WebRTC-Aec3DownSamplingFactor8KillSwitch",
+    "WebRTC-Aec3AdaptErleOnLowRenderKillSwitch",
+    "WebRTC-Aec3AgcGainChangeResponseKillSwitch",
+    "WebRTC-Aec3BoundedNearendKillSwitch",
+    "WebRTC-Aec3EarlyShadowFilterJumpstartKillSwitch",
+    "WebRTC-Aec3EnableAdaptiveEchoReverbEstimation",
+    "WebRTC-Aec3EnableLegacyDominantNearend",
+    "WebRTC-Aec3EnableUnityInitialRampupGain",
+    "WebRTC-Aec3EnableUnityNonZeroRampupGain",
     "WebRTC-Aec3EnforceSkewHysteresis1",
     "WebRTC-Aec3EnforceSkewHysteresis2",
-    "WebRTC-Aec3NewSuppressionKillSwitch",
-    "WebRTC-Aec3LinearModeWithDivergedFilterKillSwitch",
+    "WebRTC-Aec3FilterAnalyzerPreprocessorKillSwitch",
     "WebRTC-Aec3MisadjustmentEstimatorKillSwitch",
+    "WebRTC-Aec3NewFilterParamsKillSwitch",
+    "WebRTC-Aec3OverrideEchoPathGainKillSwitch",
     "WebRTC-Aec3RapidAgcGainRecoveryKillSwitch",
-    "WebRTC-Aec3SlowFilterAdaptationKillSwitch",
-    "WebRTC-Aec3SmoothUpdatesTailFreqRespKillSwitch",
-    "WebRTC-Aec3SuppressorNearendAveragingKillSwitch",
-    "WebRTC-Aec3AgcGainChangeResponseKillSwitch",
+    "WebRTC-Aec3ResetErleAtGainChangesKillSwitch",
+    "WebRTC-Aec3ReverbBasedOnRenderKillSwitch",
+    "WebRTC-Aec3ReverbModellingKillSwitch",
+    "WebRTC-Aec3ShadowFilterBoostedJumpstartKillSwitch",
     "WebRTC-Aec3ShadowFilterJumpstartKillSwitch",
-    "WebRTC-Aec3EarlyLinearFilterUsageKillSwitch",
-    "WebRTC-Aec3ShortInitialStateKillSwitch",
+    "WebRTC-Aec3ShortReverbKillSwitch",
+    "WebRTC-Aec3SmoothSignalTransitionsKillSwitch",
+    "WebRTC-Aec3SmoothUpdatesTailFreqRespKillSwitch",
+    "WebRTC-Aec3SoftTransparentModeKillSwitch",
     "WebRTC-Aec3StandardNonlinearReverbModelKillSwitch",
-    "WebRTC-Aec3EnableAdaptiveEchoReverbEstimation"};
+    "WebRTC-Aec3StrictDivergenceCheckKillSwitch",
+    "WebRTC-Aec3UseLegacyNormalSuppressorTuning",
+    "WebRTC-Aec3UseOffsetBlocks",
+    "WebRTC-Aec3UseShortDelayEstimatorWindow",
+    "WebRTC-Aec3UseStationarityPropertiesKillSwitch",
+    "WebRTC-Aec3UtilizeShadowFilterOutputKillSwitch",
+    "WebRTC-Aec3ZeroExternalDelayHeadroomKillSwitch",
+};
 
 std::unique_ptr<AudioProcessing> CreateApm(test::FuzzDataHelper* fuzz_data,
                                            std::string* field_trial_string) {