Increased high frequency transparency

Avoid excessive echo suppression in frequencies above 2 kHz when
there is a dominant nearend. Calls with clock drift will not be affected
by this change as they tend to have less accurate linear filters.

Bug: webrtc:11985
Change-Id: Iddc628da5e2ba572c1b47acd87dd3be35260dca1
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/188580
Reviewed-by: Per Åhgren <peah@webrtc.org>
Commit-Queue: Gustaf Ullberg <gustaf@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32533}

api/audio/echo_canceller3_config.h

@@ -221,6 +221,7 @@ struct RTC_EXPORT EchoCanceller3Config {
     } high_bands_suppression;
 
     float floor_first_increase = 0.00001f;
+    bool conservative_hf_suppression = false;
   } suppressor;
 };
 }  // namespace webrtc

api/audio/echo_canceller3_config_json.cc

@@ -383,6 +383,8 @@ void Aec3ConfigFromJsonString(absl::string_view json_string,
 
     ReadParam(section, "floor_first_increase",
               &cfg.suppressor.floor_first_increase);
+    ReadParam(section, "conservative_hf_suppression",
+              &cfg.suppressor.conservative_hf_suppression);
   }
 }
 
@@ -676,7 +678,10 @@ std::string Aec3ConfigToJsonString(const EchoCanceller3Config& config) {
   ost << "\"anti_howling_gain\": "
       << config.suppressor.high_bands_suppression.anti_howling_gain;
   ost << "},";
-  ost << "\"floor_first_increase\": " << config.suppressor.floor_first_increase;
+  ost << "\"floor_first_increase\": " << config.suppressor.floor_first_increase
+      << ",";
+  ost << "\"conservative_hf_suppression\": "
+      << config.suppressor.conservative_hf_suppression;
   ost << "}";
   ost << "}";
   ost << "}";

modules/audio_processing/aec3/echo_canceller3.cc

@@ -368,6 +368,10 @@ EchoCanceller3Config AdjustConfig(const EchoCanceller3Config& config) {
     adjusted_cfg.suppressor.nearend_tuning.max_dec_factor_lf = .2f;
   }
 
+  if (field_trial::IsEnabled("WebRTC-Aec3EnforceConservativeHfSuppression")) {
+    adjusted_cfg.suppressor.conservative_hf_suppression = true;
+  }
+
   if (field_trial::IsEnabled("WebRTC-Aec3EnforceStationarityProperties")) {
     adjusted_cfg.echo_audibility.use_stationarity_properties = true;
   }

modules/audio_processing/aec3/echo_remover.cc

@@ -414,12 +414,16 @@ void EchoRemoverImpl::ProcessCapture(
   const auto& echo_spectrum =
       aec_state_.UsableLinearEstimate() ? S2_linear : R2;
 
+  // Determine if the suppressor should assume clock drift.
+  const bool clock_drift = config_.echo_removal_control.has_clock_drift ||
+                           echo_path_variability.clock_drift;
+
   // Compute preferred gains.
   float high_bands_gain;
   std::array<float, kFftLengthBy2Plus1> G;
   suppression_gain_.GetGain(nearend_spectrum, echo_spectrum, R2,
                             cng_.NoiseSpectrum(), render_signal_analyzer_,
-                            aec_state_, x, &high_bands_gain, &G);
+                            aec_state_, x, clock_drift, &high_bands_gain, &G);
 
   suppression_filter_.ApplyGain(comfort_noise, high_band_comfort_noise, G,
                                 high_bands_gain, Y_fft, y);

modules/audio_processing/aec3/suppression_gain.cc

@@ -27,25 +27,24 @@
 namespace webrtc {
 namespace {
 
-void PostprocessGains(std::array<float, kFftLengthBy2Plus1>* gain) {
+void LimitLowFrequencyGains(std::array<float, kFftLengthBy2Plus1>* gain) {
-  // TODO(gustaf): Investigate if this can be relaxed to achieve higher
-  // transparency above 2 kHz.
-
   // Limit the low frequency gains to avoid the impact of the high-pass filter
   // on the lower-frequency gain influencing the overall achieved gain.
   (*gain)[0] = (*gain)[1] = std::min((*gain)[1], (*gain)[2]);
+}
 
-  // Limit the high frequency gains to avoid the impact of the anti-aliasing
+void LimitHighFrequencyGains(bool conservative_hf_suppression,
-  // filter on the upper-frequency gains influencing the overall achieved
+                             std::array<float, kFftLengthBy2Plus1>* gain) {
-  // gain. TODO(peah): Update this when new anti-aliasing filters are
+  // Limit the high frequency gains to avoid echo leakage due to an imperfect
-  // implemented.
+  // filter.
-  constexpr size_t kAntiAliasingImpactLimit = (64 * 2000) / 8000;
+  constexpr size_t kFirstBandToLimit = (64 * 2000) / 8000;
-  const float min_upper_gain = (*gain)[kAntiAliasingImpactLimit];
+  const float min_upper_gain = (*gain)[kFirstBandToLimit];
   std::for_each(
-      gain->begin() + kAntiAliasingImpactLimit, gain->end() - 1,
+      gain->begin() + kFirstBandToLimit + 1, gain->end(),
       [min_upper_gain](float& a) { a = std::min(a, min_upper_gain); });
   (*gain)[kFftLengthBy2] = (*gain)[kFftLengthBy2Minus1];
 
+  if (conservative_hf_suppression) {
     // Limits the gain in the frequencies for which the adaptive filter has not
     // converged.
     // TODO(peah): Make adaptive to take the actual filter error into account.
@@ -58,8 +57,10 @@ void PostprocessGains(std::array<float, kFftLengthBy2Plus1>* gain) {
                         gain->begin() + kUpperAccurateBandPlus1, 0.f) *
         oneByBandsInSum;
 
-  std::for_each(gain->begin() + kUpperAccurateBandPlus1, gain->end(),
+    std::for_each(
+        gain->begin() + kUpperAccurateBandPlus1, gain->end(),
         [hf_gain_bound](float& a) { a = std::min(a, hf_gain_bound); });
+  }
 }
 
 // Scales the echo according to assessed audibility at the other end.
@@ -265,6 +266,7 @@ void SuppressionGain::LowerBandGain(
         suppressor_input,
     rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> residual_echo,
     rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> comfort_noise,
+    bool clock_drift,
     std::array<float, kFftLengthBy2Plus1>* gain) {
   gain->fill(1.f);
   const bool saturated_echo = aec_state.SaturatedEcho();
@@ -298,8 +300,14 @@ void SuppressionGain::LowerBandGain(
               last_echo_[ch].begin());
   }
 
-  // Limit high-frequency gains.
+  LimitLowFrequencyGains(gain);
-  PostprocessGains(gain);
+  // Use conservative high-frequency gains during clock-drift or when not in
+  // dominant nearend.
+  if (!dominant_nearend_detector_->IsNearendState() || clock_drift ||
+      config_.suppressor.conservative_hf_suppression) {
+    LimitHighFrequencyGains(config_.suppressor.conservative_hf_suppression,
+                            gain);
+  }
 
   // Store computed gains.
   std::copy(gain->begin(), gain->end(), last_gain_.begin());
@@ -352,6 +360,7 @@ void SuppressionGain::GetGain(
     const RenderSignalAnalyzer& render_signal_analyzer,
     const AecState& aec_state,
     const std::vector<std::vector<std::vector<float>>>& render,
+    bool clock_drift,
     float* high_bands_gain,
     std::array<float, kFftLengthBy2Plus1>* low_band_gain) {
   RTC_DCHECK(high_bands_gain);
@@ -364,7 +373,8 @@ void SuppressionGain::GetGain(
   // Compute gain for the lower band.
   bool low_noise_render = low_render_detector_.Detect(render);
   LowerBandGain(low_noise_render, aec_state, nearend_spectrum,
-                residual_echo_spectrum, comfort_noise_spectrum, low_band_gain);
+                residual_echo_spectrum, comfort_noise_spectrum, clock_drift,
+                low_band_gain);
 
   // Compute the gain for the upper bands.
   const absl::optional<int> narrow_peak_band =

modules/audio_processing/aec3/suppression_gain.h

@@ -47,6 +47,7 @@ class SuppressionGain {
       const RenderSignalAnalyzer& render_signal_analyzer,
       const AecState& aec_state,
       const std::vector<std::vector<std::vector<float>>>& render,
+      bool clock_drift,
       float* high_bands_gain,
       std::array<float, kFftLengthBy2Plus1>* low_band_gain);
 
@@ -76,6 +77,7 @@ class SuppressionGain {
           suppressor_input,
       rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> residual_echo,
       rtc::ArrayView<const std::array<float, kFftLengthBy2Plus1>> comfort_noise,
+      bool clock_drift,
       std::array<float, kFftLengthBy2Plus1>* gain);
 
   void GetMinGain(rtc::ArrayView<const float> weighted_residual_echo,

modules/audio_processing/aec3/suppression_gain_unittest.cc

@@ -49,7 +49,7 @@ TEST(SuppressionGainDeathTest, NullOutputGains) {
                    std::vector<std::vector<std::vector<float>>>(
                        3, std::vector<std::vector<float>>(
                               1, std::vector<float>(kBlockSize, 0.f))),
-                   &high_bands_gain, nullptr),
+                   false, &high_bands_gain, nullptr),
       "");
 }
 
@@ -107,7 +107,7 @@ TEST(SuppressionGain, BasicGainComputation) {
     aec_state.Update(delay_estimate, subtractor.FilterFrequencyResponses(),
                      subtractor.FilterImpulseResponses(),
                      *render_delay_buffer->GetRenderBuffer(), E2, Y2, output);
-    suppression_gain.GetGain(E2, S2, R2, N2, analyzer, aec_state, x,
+    suppression_gain.GetGain(E2, S2, R2, N2, analyzer, aec_state, x, false,
                              &high_bands_gain, &g);
   }
   std::for_each(g.begin(), g.end(),
@@ -126,7 +126,7 @@ TEST(SuppressionGain, BasicGainComputation) {
     aec_state.Update(delay_estimate, subtractor.FilterFrequencyResponses(),
                      subtractor.FilterImpulseResponses(),
                      *render_delay_buffer->GetRenderBuffer(), E2, Y2, output);
-    suppression_gain.GetGain(E2, S2, R2, N2, analyzer, aec_state, x,
+    suppression_gain.GetGain(E2, S2, R2, N2, analyzer, aec_state, x, false,
                              &high_bands_gain, &g);
   }
   std::for_each(g.begin(), g.end(),
@@ -137,7 +137,7 @@ TEST(SuppressionGain, BasicGainComputation) {
   R2[1].fill(10000000000000.f);
 
   for (int k = 0; k < 10; ++k) {
-    suppression_gain.GetGain(E2, S2, R2, N2, analyzer, aec_state, x,
+    suppression_gain.GetGain(E2, S2, R2, N2, analyzer, aec_state, x, false,
                              &high_bands_gain, &g);
   }
   std::for_each(g.begin(), g.end(),

resources/audio_processing/output_data_fixed.pb.sha1

@@ -1 +1 @@
-4010b1fe15eda1b42968cdb3f9fed399e1aa7197
+0ff9ab4d46929552e21d16f266f9eba42575ba8d

resources/audio_processing/output_data_float.pb.sha1

@@ -1 +1 @@
-d22d4b0bc8f59aa27da61e158b9d35596f3844f5
+ed1172c80a1a001a8aa7ac0680a99018cbb7d278

resources/audio_processing/output_data_float_avx2.pb.sha1

@@ -1 +1 @@
-514543fbee78d0a71e87adb92e23138d762d1da8
+a1dd718a6882bf8033a934e5beec73086cc91240