On Fri, 2012-03-16 at 17:34 +0000, James Lee wrote: > A/V sync is handled in the decoders based on the current PCR and PTS values for > each audio/video frame. Our concern is audio latency through pulse, not A/V > synchronization. The latency jumps right after the gap. I'd still say that your problem is A/V sync, not latency, unless the latency information from Pulseaudio is incorrect. Doing A/V sync requires you to know how long it will take for the audio to come out of the speakers and for the video frames to be visible if you write data now, and make timing decisions based on that. If you do your timing decisions only based on the PCR and PTS values, it sounds like you expect to hear the audio immediately when you issue a write command in your code. That said, it sounds like the assumption works well enough for you when you play the first movie. I don't really see a reason why continuing after a stream underrun should cause any longer delays than starting a new stream, unless there's prebuffering involved in the underrun case. If there's prebuffering in the underrun case, I'd expect there to be similar prebuffering also in the stream creation case, though... I guess you have checked that the A/V sync is actually still ok when the previous movie ends, and it only goes wrong when the next one starts? If you don't resample to compensate the drift between the server clock and the sound card clock, data can accumulate slowly in the stream buffer. Then again, stopping writing data during the gap between the movies should drain any accumulated data and reset the situation... I guess this was not very helpful in finding a solution. It sounds like there may be a bug in Pulseaudio or the alsa plugin. You can file a bug if you wish. A test program that shows that the latency is misbehaving would be great. Assuming that Pulseaudio reports the latency correctly, it should be also possible for you to fix your application to monitor the latency and adapt to jumps as needed. Also recreating the stream after the gap could be a simple workaround? -- Tanu
