Opened 5 years ago

Last modified 5 years ago

#11163 new bug

Audio/Video out of sync with some .ogv files

Reported by: humdinger Owned by: colin
Priority: normal Milestone: R1
Component: Kits/Media Kit Version: R1/Development
Keywords: Cc:
Blocked By: Blocking: #6976
Has a Patch: no Platform: All

Description

This is hrev47733.

Some .ogv files show non-snc audio/video, e.g. in mmu_man's screen-capture [40 MiB] the audio is about 35 secs late when played with MediaPlayer.

Other .ogv do not show this, e.g. Bug Bunny trailers.

Might be the different encodings:
RMLL: 15 fps, 22.050 kHz, 90 kbps
Bunny: 25 fps, 48 kHz, 80 kbps

Change History (2)

comment:1 by tangobravo, 5 years ago

This sounds very similar to an issue I tracked down at work a couple of years ago with Theora video sync. The issue arose in videos that contained duplicate frames - something that is probably quite common with screen-capture videos.

Page 17 of the Theora spec at http://www.theora.org/doc/Theora.pdf mentions that Theora can code duplicate frames using empty (0-byte) packets. Not all Theora encoders do this, but some definitely do (slightly hazy memory but ffmpeg2theora, which uses the reference encoder, did produce these 0-byte duplicate packets, whereas a straight ffmpeg conversion to ogg did not).

The ogg container format doesn't specify the individual frame numbers for each packet, the container only specifies the last one in each page. I think that means that ffmpeg doesn't really use the frame number in the container other than for seeking, but just increments it for every packet. I think the cause of the bug was then down to the fact ffmpeg did not handle these "0-byte packets" correctly as being duplicate frames, but thought they were corrupt data.

Again, hazy memory, but it was either just not counting them as frames, so the video was all played too quickly, or it was correctly counting them but thinking they were corrupt and thus jumping past all of the duplicates and displaying the next (non-duplicate frame) far too early (but keeping them on screen so long-term drift didn't get worse).

One of the ways I got to these conclusions was by getting ffmpeg to try to round-trip a PNG sequence to theora ogg and back. Also try the encode with ffmpeg2theora and back with ffmpeg.

A sequence AAAAABBBBBCCCCCDDDDD (where A, B, C, D are each a different PNG files, repeated 5 times in that sequence) should prove informative. I think I actually did something with a script to make 25 copies of each png, and use a 1-second-beep audio loop to debug this. There were definite differences between firefox/vlc/ffplay in terms of how well they dealt with those files.

Humdinger: Does sync recover after a seek? Is this new behaviour after Colin's recent media changes?

Last edited 5 years ago by tangobravo (previous) (diff)

comment:2 by pulkomandy, 5 years ago

Blocking: 6976 added
Note: See TracTickets for help on using tickets.