[janlauber]

Voices sounded over-cancelled while music, the radio or a video played. Noise cancellation was not the cause: Chrome's echo canceller was.

echoCancellation: true is Chrome-wide echo cancellation. Its reference is the output device, so music counted as echo, and the canceller suppressed the rider's voice under it, on headphones too.

The change

• Capture asks for echoCancellation: 'remote-only', so only audio an element plays counts as echo.
• Voices reach the speakers through WebAudio, which remote-only does not see. Each voice's element now plays a copy at ECHO_REFERENCE (0.2, 14 dB under the graph), on the same output. It goes to 0 at fader 0 and while Away.
• A share stays out, because it is music. A handheld and a mic test keep true.

createMediaElementSource was the obvious way to avoid a second copy. It does not work: WebMediaPlayerMS has no audio provider, so the element keeps playing.

Numbers

Raw capture through the real model and gate, speech at −26 dBFS. With --mute-audio, true cancels nothing, so the "before" rows put music in the same canceller's reference through a remote track under remote-only.

Case	PESQ-WB	STOI	First 80 ms of a word
No music	4.40	0.999	−0.3 dB
Headphones, music in the reference	1.58	0.815	−7.2
Headphones, music 12 dB quieter	3.02	0.944	−4.4
Speakers at −30 dBFS, music in the reference	1.24	0.516	−14.4
Speakers at −30 dBFS, music out of it	3.06	0.973	0.1
Speakers, a song in the reference	1.37	0.702	−47.7
Speakers, a song out of it	1.38	0.852	3.8

Crew echo in a real call through LiveKit: a second rider's mic carries the first rider's voice at −26 dBFS. Without the copy it goes back at −24 dBFS. With the copy it reaches −56 dBFS within 10 s and stays between −56 and −81. The element-volume sweep is in docs/RESEARCH.md §22.6: 0.2 cancels, 0.1 cancels nothing.

Regression, real call, at the listener against #4433: quiet 4.04 → 4.07, fan 3.19 → 3.20, bike 2.72 → 2.71. One quiet run scored 3.22 with 0.58 % concealment. Two reruns scored 4.07 with none, so I read that one as a stall.

What it costs

• A song on speakers now leaks into the call at about −31 dBFS, because a singer is speech to the model. Instrumental music does not.
• Each voice has a second copy 14 dB down, which ripples its level by about 2 dB at most.
• A mic test keeps true, so it can sound worse over music than the call does.

Not verified

• true with music on real speakers. Reproducing it needs sound out of the Mac, so the before rows use the same canceller by another route.
• How the copy sounds. Nobody has listened to it, and its delay against the graph is unmeasured.
• A rider talking over the crew on speakers.