To implement a full AEC for getUserMedia, where all channels are cancelled along with other browser-generated audio, we need access to the full, final output stream as close to where it goes to the OS/speakers as possible, including all underrun/overrun modifications. It will likely be in the output clock domain, and input devices will be a different clock domain. The AEC will need to compensate for that (as normal). We'll also want to know about any deeper-than-where-we-get-this-data underruns or overruns that will affect timing. If we submitted 10ms of data to the lower levels (and reported it to the AEC/gUM), and then we're told there was an underrun down lower, we likely will need to reset/retrain the AEC. Also notification (if possible) if the output device changes (speakers to headphones, etc) if we know. Note that we may want such a signal to be async to the buffers being forwarded. We'll want to use a lockless circular buffer for the audio data, likely, if possible. (there's one in the mac webrtc driver IIRC)

Fixed by 982490