Philipp Hancke tests why voice agents sometimes answer their own greetings. In his setup, echo reached transcription in 21% of Edge sessions and 4% of Chrome sessions. AEC dumps show how echo cancellation, gain control, and VAD thresholds interact.
Technology
Google Chrome downloads a new neural AEC while you are not looking
Chrome quietly delivers a neural model that changes part of its echo-cancellation pipeline. Philipp Hancke reverse engineers the model and compares echo suppression and speech recognition against AEC3.
Comparing WebRTC video codecs with a proper dataset
Stumbling over Microsoft’s Video Conferencing Dataset (VCD) I asked Claude to take on a video codec engineer’s favorite hot take: compare video codecs. It did a fairly good writeup using the tooling provided by libWebRTC. One of the most surprising results was that background blur makes H.264 a lot more effective. Other than that, the […]
Vector Vibing to speed up Opus encode by 20%
tl;dr: Opus encoding just got 20% faster on desktop Chromium by turning on AVX2 vectorization in libopus. Vectorization means one instruction doing the same arithmetic on a whole vector of values at once instead of one at a time. Codecs spend their hot loops running identical math across long buffers of samples, exactly where vectorization […]
WebRTC vs. MoQ by Use Case
This might be my first editorial style post here. Fippo’s Is everyone switching to MoQ from WebRTC? started some threads on MoQ vs. WebRTC. I started to respond, but my responses quickly became too long so I decided to go even deeper with a post here. Fippo’s post shows hard data that Media over QUIC […]