I'm writing simple HLS (Http Live Streaming) java server to live cast (really live, not on demand) screenshow + voice. I constantly get chunks of image frames and audio samples as input to my service and produce mpeg 2 ts files + m3u8 playlist web page as output. The workflow is the following:
- Collect (buffer) source video frames and audio for certain period of time
- Convert series of video frames to h.264 encoded video file
- Convert audio samples to mp3 audio file
Merge them to
.ts
file with ffmpeg commandffmpeg -i audio.mp3 -i video.mp4 -f mpegts -c:a copy -c:v copy -vprofile main -level:v 4.0 -vbsf h264_mp4toannexb -flags -global_header segment.ts
Publish several
.ts
files on m3u8 playlist.
The problem is resulting playlist interrupts after first segment is played. VLC logs following error:
freetype error: Breaking unbreakable line
ts error: libdvbpsi (PSI decoder): TS discontinuity (received 0, expected 4) for PID 17
ts error: libdvbpsi (PSI decoder): TS duplicate (received 0, expected 1) for PID 0
ts error: libdvbpsi (PSI decoder): TS duplicate (received 0, expected 1) for PID 4096
core error: ES_OUT_SET_(GROUP_)PCR is called too late (pts_delay increased to 1000 ms)
core error: ES_OUT_RESET_PCR called
core error: Could not convert timestamp 185529572000
ts error: libdvbpsi (PSI decoder): TS discontinuity (received 0, expected 4) for PID 17
ts error: libdvbpsi (PSI decoder): TS duplicate (received 0, expected 1) for PID 0
ts error: libdvbpsi (PSI decoder): TS duplicate (received 0, expected 1) for PID 4096
core error: ES_OUT_SET_(GROUP_)PCR is called too late (jitter of 8653 ms ignored)
core error: Could not get display date for timestamp 0
core error: Could not convert timestamp 185538017000
core error: Could not convert timestamp 185538267000
core error: Could not convert timestamp 185539295977
...
I guess the reason is that start time of segments do not belong to one stream, but it's impossible to concat and resegment (with ffmepg -f segment
) whole stream once new chunk is added. Tried adding #EXT-X-DISCONTINUITY
tag to playlist as suggested here but it didn't help. When I ffprobe
them I get:
Input #0, mpegts, from '26.ts':
Duration: 00:00:10.02, start: 1.876978, bitrate: 105 kb/s
Program 1
Metadata:
service_name : Service01
service_provider: FFmpeg
Stream #0:0[0x100]: Video: h264 (High) ([27][0][0][0] / 0x001B), yuv420p, 640x640, 4 fps, 4 tbr, 90k tbn, 8 tbc
Stream #0:1[0x101]: Audio: mp3 ([3][0][0][0] / 0x0003), 48000 Hz, mono, s16p, 64 kb/s
Where start value in line Duration: 00:00:10.02, start: 1.876978, bitrate: 105 kb/s
is more or less equal for all segments.
When I check segments from available proven-to-work playlists (like http://vevoplaylist-live.hls.adaptive.level3.net/vevo/ch1/appleman.m3u8) they all have diffrenet start values for each segment, for example:
Input #0, mpegts, from 'segm150518140104572-424570.ts':
Duration: 00:00:06.17, start: 65884.808689, bitrate: 479 kb/s
Program 257
Stream #0:0[0x20]: Video: h264 (Constrained Baseline) ([27][0][0][0] / 0x001B), yuv420p, 320x180 [SAR 1:1 DAR 16:9], 30 fps, 29.97 tbr, 90k tbn, 60 tbc
Stream #0:1[0x21]: Audio: aac (LC) ([15][0][0][0] / 0x000F), 44100 Hz, stereo, fltp, 115 kb/s
Stream #0:2[0x22]: Data: timed_id3 (ID3 / 0x20334449)
and the next after it
Input #0, mpegts, from 'segm150518140104572-424571.ts':
Duration: 00:00:06.22, start: 65890.814689, bitrate: 468 kb/s
Program 257
Stream #0:0[0x20]: Video: h264 (Constrained Baseline) ([27][0][0][0] / 0x001B), yuv420p, 320x180 [SAR 1:1 DAR 16:9], 30 fps, 29.97 tbr, 90k tbn, 60 tbc
Stream #0:1[0x21]: Audio: aac (LC) ([15][0][0][0] / 0x000F), 44100 Hz, stereo, fltp, 124 kb/s
Stream #0:2[0x22]: Data: timed_id3 (ID3 / 0x20334449)
differ in the way that start time of segm150518140104572-424571.ts
is equal to start time + duration of segm150518140104572-424570.ts
.
How could this start value be adjusted with ffmpeg
? Or maybe my whole aproach is wrong? Unfortunately I couldn't find on the internet working example of live (not on demand) video service implemented with ffmepg.
I've answered my own question with approach described here. Timestamp could be adjusted with
-initial_offset
parameter of-f segment
format:For each new segment I calculate necessary
-initial_offset
as sum of all previous segments length.