From 85a11a35f176927cddc82a6a2af9072a338cb9e7 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Per=20=C3=85hgren?= <peah@webrtc.org>
Date: Mon, 2 Oct 2017 14:42:06 +0200
Subject: [PATCH] Bounding the AEC3 suppression gain for poorly estimated
 residual echoes
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

This CL bounds the supppression gain for higher frequencies where
the estimate of the residual echo sometimes is less accurate.

Bug: webrtc:8320
Change-Id: I02b21e6b1758c7e8b6660c1631a05c956a45e4c8
Reviewed-on: https://webrtc-review.googlesource.com/5260
Reviewed-by: Gustaf Ullberg <gustaf@webrtc.org>
Commit-Queue: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#20081}
---
 .../audio_processing/aec3/suppression_gain.cc | 28 ++++++++++++++++++-
 1 file changed, 27 insertions(+), 1 deletion(-)

diff --git a/modules/audio_processing/aec3/suppression_gain.cc b/modules/audio_processing/aec3/suppression_gain.cc
index 0bed664584..83ae63c95e 100644
--- a/modules/audio_processing/aec3/suppression_gain.cc
+++ b/modules/audio_processing/aec3/suppression_gain.cc
@@ -192,6 +192,9 @@ void GainToNoAudibleEcho(
   }
 }
 
+// TODO(peah): Make adaptive to take the actual filter error into account.
+constexpr size_t kUpperAccurateBandPlus1 = 29;
+
 // Computes the signal output power that masks the echo signal.
 void MaskingPower(const AudioProcessing::Config::EchoCanceller3& config,
                   const std::array<float, kFftLengthBy2Plus1>& nearend,
@@ -205,11 +208,31 @@ void MaskingPower(const AudioProcessing::Config::EchoCanceller3& config,
     (*masker)[k] =
         comfort_noise[k] + config.param.gain_mask.m4 * last_masker[k];
   }
-  for (size_t k = 1; k < gain.size() - 1; ++k) {
+
+  // Apply masking only between lower frequency bands.
+  RTC_DCHECK_LT(kUpperAccurateBandPlus1, gain.size());
+  for (size_t k = 1; k < kUpperAccurateBandPlus1; ++k) {
     (*masker)[k] += 0.1f * (side_band_masker[k - 1] + side_band_masker[k + 1]);
   }
 }
 
+// Limits the gain in the frequencies for which the adaptive filter has not
+// converged. Currently, these frequencies are not hardcoded to the frequencies
+// which are typically not excited by speech.
+// TODO(peah): Make adaptive to take the actual filter error into account.
+void AdjustNonConvergedFrequencies(
+    std::array<float, kFftLengthBy2Plus1>* gain) {
+  constexpr float oneByBandsInSum =
+      1 / static_cast<float>(kUpperAccurateBandPlus1 - 20);
+  const float hf_gain_bound =
+      std::accumulate(gain->begin() + 20,
+                      gain->begin() + kUpperAccurateBandPlus1, 0.f) *
+      oneByBandsInSum;
+
+  std::for_each(gain->begin() + kUpperAccurateBandPlus1, gain->end(),
+                [hf_gain_bound](float& a) { a = std::min(a, hf_gain_bound); });
+}
+
 }  // namespace
 
 // TODO(peah): Add further optimizations, in particular for the divisions.
@@ -270,6 +293,9 @@ void SuppressionGain::LowerBandGain(
     }
   }
 
+  // Adjust the gain for frequencies which have not yet converged.
+  AdjustNonConvergedFrequencies(gain);
+
   // Update the allowed maximum gain increase.
   UpdateMaxGainIncrease(config_, no_saturation_counter_, low_noise_render,
                         last_echo_, echo, last_gain_, *gain, &gain_increase_);