Re-tuning of VAD in AGC2.
Changing VAD (voice activity detector) confidence threshold from 40%
to 90%. The proportion of samples classified as speech drops to ca 80%
of what it was when the threshold was 40%. Therefore,
kFullBufferSizeMs has to be increased by 1.0/0.8. We increase it from
1600ms to 2000ms.
TESTED = Did run the new and old configs on AEC dumps. With one minute
of kitchen noise, the new tuning boosted the noise by 3-4 db less.
Bug: chromium:913430
Change-Id: I4a2ebb6d1d309c6c20dd23c3685818b1b5ad4a66
Reviewed-on: https://webrtc-review.googlesource.com/c/113806
Commit-Queue: Alex Loiko <aleloi@webrtc.org>
Reviewed-by: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#25950}diff --git a/modules/audio_processing/agc2/agc2_common.h b/modules/audio_processing/agc2/agc2_common.h
index 55dd648..a6389f4 100644
--- a/modules/audio_processing/agc2/agc2_common.h
+++ b/modules/audio_processing/agc2/agc2_common.h
@@ -41,10 +41,10 @@
// This is the threshold for speech. Speech frames are used for updating the
// speech level, measuring the amount of speech, and decide when to allow target
// gain reduction.
-constexpr float kVadConfidenceThreshold = 0.4f;
+constexpr float kVadConfidenceThreshold = 0.9f;
// The amount of 'memory' of the Level Estimator. Decides leak factors.
-constexpr size_t kFullBufferSizeMs = 1600;
+constexpr size_t kFullBufferSizeMs = 1200;
constexpr float kFullBufferLeakFactor = 1.f - 1.f / kFullBufferSizeMs;
constexpr float kInitialSpeechLevelEstimateDbfs = -30.f;
