[FFmpeg-devel] Matroska and output_ts_offset

[Andreas Rheinhardt]

Peter Zebühr:
> Hello,
> 
> I noticed the other day that if I try to mux opus in webm with a timestamp offset set I don't get the expected timestamps out on my packets. Example:
> 
> 	ffmpeg -f lavfi -i "sine=frequency=1000:duration=60" -output_ts_offset 200ms -c:a libopus sine_200ms_offset.webm
> 
> Now if I inspect the packets with ffprobe on the produced output the first packet has a PTS of 186. My expectation would be that it would be: 200 - 6.5 (opus encoder delay) ~= 193.
> Looking at the output with mkvinfo also indicates that it starts at 193ms (186 + encoder delay I assume here).
> 
> At first I thought it looked like the encoder delay gets added twice, but after some more pondering I think what is happening here is:
> 
> 1. Opus outputs packets starting with the priming samples, that ends up with a negative PTS (rounded to -7)
> 2. In mux.c the output_ts_offset takes effect before the call to "handle_avoid_negative_ts"
> 3. PTS -7 gets shifter by the requested ts_offset to 193
> 4. Matroska muxer writes the side data about encoder delay. 
> 5. The produced stream now effectively starts one encoder delay distance too early.
> 
> I played around locally with moving the call to "handle_avoid_negative_ts" in mux.c/write_packet to right before the ts_offset handling instead of right after and that seems to fix this particular problem. But, I am unsure what other potential consequences that would have. Would appreciate if someone more familiar with this can help fix this issue.
> 

The Matroska muxer is buggy wrt. to the ts_offset relating to codec
delay. You can see it in lines 1839-1841 which are commented out.
Commenting them out happened in commit
82e4f39883932c1b1e5c7792a1be12dec6ab603d, merging the libav commit that
implemented it (namely
https://github.com/FFmpeg/FFmpeg/commit/a1aa37dd0b96710d4a17718198a3f56aea2040c1).
It mentions "assertion failures and av sync errors". I can only think of
one way that it could have led to assertion failures, but this should
have been fixed in 4ebeab15b037a21f195696cef1f7522daf42f3ee (and since
then I wondered whether it can't be enabled).

- Andreas