Each speaker's stream held a fixed 60 ms and gave up after 200 ms of
concealment, so a jittery connection concealed constantly and a stall
ended the burst. It now runs through a port of the libspeex jitter
buffer the official client links, configured and driven as it does:
- the delay adapts to the measured jitter, aiming for at most 1% of
packets arriving late
- a speaker keeps one timeline, never re-anchored between talk bursts,
so what the buffer learned carries over; only 5 s without packets
drops it
- the talker's stop marker never enters the buffer; playback ends when
it reaches it, and its slot is played as silence, not concealed
A pull now always hands out one 20 ms frame: what a longer packet or a
delay increase produces beyond that is played on the following pulls,
which Android's mixer used to cut off.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ericek111
0083dc8ad6