[audio perception] Merge voice streams

[avaer]

In Discord voice we were relying on the client start/stops, but they are too noisy.

If we merge the streams per user (with inserted silence) and use our own VAD (which is pretty good), I think the results would be better.