I create a webrtc peer connection with my server(only stun)
Using pion webrtc for the server
I write the received RTP packets as VP8 and opus streams, as described here, to two pipes (the writers; created with os.Pipe() in golang)
The read ends of these two pipes are received by ffmpeg as inputs (via exec.Command.ExtraFiles) for transcoding using libx264 and aac into a single stream. The command:
ffmpeg -re -i pipe:3 -re -r pipe:4 -c:a aac -af aresample=48000 -c:v libx264 -x264-params keyint=48:min-keyint=24 -profile:v main -preset ultrafast -tune zerolatency -crf 20 -fflags genpts -avoid_negative_ts make_zero -vsync vfr -map 0:0,0:0 -map 1:0,0:0 -f matroska -strict -2 pipe:5
The above command outputs to a pipe(:5) the read end of which is being taken as input by the following:
ffmpeg -hide_banner -y -re -i pipe:3 -sn -vf scale=-1:'min(ih,360)' -c:v libx264 -pix_fmt yuv420p -ca aac -b:a 128k -b:v 1400k -maxrate 1498k -bufsize 2100k -hls_time 1 -hls_playlist_type event -hls_base_url /workdir/streamID/360p -hls_segment_filename /workdir/streamID/360p/360_%%03d.ts -f hls /workdir/streamID/360p.m3u8
This works fine as long as there are no movements of my webcam. The moment that happens the video speed suddenly increases for a split second and audio delay gets introduced. This delay keeps increasing each time I shake my webcam.
The first command in point 4 above - if written to a file separately will be absolutely fine, in terms of a/v sync, even with vigorous camera shaking. The weird audio delay is only when transcoding for hls output irrespective of whether I'm actually viewing it live or playing it back later.
This is my first time working with ffmpeg/hls/webrtc - would be really helpful if I could be pointed in the correct direction at least to be able to debug this or even know why this happens. Any and all help is greatly appreciated
I'm doing screen recordings of gameplay (Dota2) using my NVIDIA graphics card GeForce experience hardware recording (NVEC Encoder). This creates a variable frame rate output video. My NVIDIA settings are 60 fps 15000 kbps. I have paid a guy to make a program that generates scripts that given start/stop timepoints can extract clips from the video and merge them with crossfade. See example code below. The script works for many input recordings but fails often: The audio and video are desynchronized (usually audio delay) in many of the clips, ca 0.5 seconds. I think it fails more when frame rate dropped more during recording. He does not know how to fix the problem, and I wonder if anyone could point out if anything could be fixed in the script (example below)?
Processing speed is quite important (now making a 10 min 'highlight' video takes ca 7-10 min). Solutions increasing that amount very much more is not of too big interest, unfortunately. His approach has been to work separately with audio and video and merge in the end. He already has a program to make ffmpeg code for working with different scenarios (also adding overlays, adding music, intro/outro) so it would be preferable with some easy fixes to his code and not dramatic redesigning of the logic. But if nothing else can fix the problem, a redesign in logic is ok. Using other tools than ffmpeg is also ok, but should be automatable (scripts/cli) and not increase processing times too much.
Running the program "mediainfo" on the input video shows that framerate dropped quite low for this input video:
Frame rate mode: Variable
Frame rate : 60.000 FPS
Minimum frame rate: 3.059 FPS
Maximum frame rate: 63.739 FPS
Full report here: https://pastebin.com/TX061Wih
The input video can be downloaded from dropbox here (6 GB):
https://www.dropbox.com/s/ftwdgapazbi62pr/fullgame.mp4?dl=0
Here the example of a script when asked to extract two clips from input video at 9:57 (41 sec length) and 15:45 (28 sec length) and crossfade merge them with a 0.5 crossfade time. There might be some code-remnants from options that are not used in this example (overlays, music, intro/outro). Using the input video above, this creates audio/video desync.
6 commands excecuted in sequence:
ffmpeg.exe -loglevel warning -ss 00:09:57 -i fullgame.mp4 -t 00:00:41 -filter_complex "[0:a]afade=t=out:st=40.5:d=0.5[a1]" -map "[a1]" -y out_temp_00.mp4.wav
ffmpeg.exe -loglevel warning -i fullgame.mp4 -ss 00:09:57 -t 00:00:41 -an -vcodec copy -f mpegts -avoid_negative_ts make_zero -y out_temp_00.mp4.ts
ffmpeg.exe -loglevel warning -ss 00:15:45 -i fullgame.mp4 -t 00:00:28 -filter_complex "[0:a]afade=t=in:st=0:d=0.5[a1]" -map "[a1]" -y out_temp_01.mp4.wav
ffmpeg.exe -loglevel warning -i fullgame.mp4 -ss 00:15:45 -t 00:00:28 -an -vcodec copy -f mpegts -avoid_negative_ts make_zero -y out_temp_01.mp4.ts
ffmpeg.exe -loglevel warning -i out_temp_00.mp4.wav -i out_temp_01.mp4.wav -y -filter_complex "[0:a]adelay=0|0[a0];[1:a]adelay=40500|40500[a1];[a0][a1]amix=inputs=2:dropout_transition=68.5,atrim=duration=68.5[outa0];[outa0]loudnorm[outa]" -map "[outa]" -ar 48000 -acodec aac -strict -2 fullgame_Output.mp4.aac
ffmpeg.exe -loglevel warning -i out_temp_00.mp4.ts -i out_temp_01.mp4.ts -y -i fullgame_Output.mp4.aac -filter_complex "[0:v]trim=start=0.5,setpts=PTS-STARTPTS[0c];[1:v]trim=start=0.5,setpts=PTS-STARTPTS[1c];[0:v]trim=40.5:41,setpts=PTS-STARTPTS[fo];[1:v]trim=0:0.5[fi];[fi]format=pix_fmts=yuva420p,fade=t=in:st=0:d=0.5:alpha=1[z];[fo]format=pix_fmts=yuva420p,fade=t=out:st=0:d=0.5:alpha=1[x];[z]fifo[w];[x]fifo[q];[q][w]overlay[r];[0c][r][1c]concat=n=3[outv]" -map "[outv]" -map 2:a -shortest -acodec copy -vcodec libx264 -preset ultrafast -b 15000k -aspect 1920:1080 fullgame_Output.mp4
P.S.
I already asked for help at an ffmpeg chat room. One guy said he knew what the problem was, but didnt know how to fix it(?):
[00:10] <kepstin> oh, wait, you're using -vcodec copy
[00:10] <kepstin> that explains everything.
[00:10] <kepstin> when you're using -vcodec copy, the start time (set with -ss) is rounded to the nearest keyframe
[00:10] <kepstin> it's not exact
[00:11] <kepstin> depending on the keyframe interval, this will result in possibly quite large shifts
[00:11] <kepstin> (also, your commands are applying audio filters on commands with -an, which is confusing/contradictory)
[00:12] <birdboy88> so the problem is that the audio temporary clips are not being extracted from the same excat timepoints?
[00:13] <kepstin> birdboy88: yeah, your audio is being re-encoded to wav so it's being cut sample-accurate, but the video's not being precisely cut.
[00:16] <birdboy88> kepstin: so I need to use slow seek (?) to extract video accurately? Or somehow extract audio only where there are video keyframes?
[00:17] <kepstin> birdboy88: i don't know how to extract audio starting at video keyframes with ffmpeg cli. You're already doing slow seek, which doesn't help (you should move the -ss option to before the -i option to speed it up)
[00:17] <kepstin> if you want accurate video cutting when saving to a file, you have to re-encode the video
[00:18] <kepstin> (doing this in a single ffmpeg command means you don't have to save to a file, so you can avoid the issue)
[00:18] * kepstin is off for a bit now
EDIT:
Everything is done with the latest ffmpeg version.
I was unable to get Gyan's code to work. It always loses some audio (audio is either 40.5 or 27.5, so only one audio is used). This is the only one working for me (changes were adelay=40500|40500 and amix=inputs=2[a0];[a0]loudnorm):
ffmpeg -i fullgame.mp4 -filter_complex "[0]split=2[vpre][vpost];
[0]asplit=2[apre][apost];
[vpre]trim=start='00:09:57':duration='00:00:41',setpts=PTS-STARTPTS[vpre-t];
[apre]atrim=start='00:09:57':duration='00:00:41',asetpts=PTS-STARTPTS,afade=t=out:st=40.5:d=0.5[apre-t];
[vpost]trim=start='00:15:45':duration='00:00:28',setpts=PTS-STARTPTS,format=yuva420p,fade=t=in:st=0:d=0.5:alpha=1,setpts=PTS+40.5/TB[vpost-t];
[apost]atrim=start='00:15:45':duration='00:00:28',asetpts=PTS-STARTPTS,afade=t=in:st=0:d=0.5,adelay=40500|40500[apost-t];
[vpre-t][vpost-t]overlay[v];
[apre-t][apost-t]amix=inputs=2[a0];[a0]loudnorm[a]" -map "[v]" -map "[a]" -y -c:v libx264 -preset ultrafast -b:v 15000k -aspect 1920:1080 -c:a aac fullgame_Output.mp4
Then I tried using a similar setup but with 3 clips, but on one machine I got error: "Error while filtering: Cannot allocate memory". And my 16 GB memory machine the processing speed is 0.02x! Any way to avoid this? This is the code I tried:
ffmpeg -i fullgame.mp4 -filter_complex "[0]split=3[vpre][vpost][v3];
[0]asplit=3[apre][apost][a3];
[vpre]trim=start=357:duration=41,setpts=PTS-STARTPTS[vpre-t];
[apre]atrim=start=357:duration=41,asetpts=PTS-STARTPTS,afade=t=out:st=40.5:d=0.5[apre-t];
[vpost]trim=start=795:duration=28,setpts=PTS-STARTPTS,format=yuva420p,fade=t=in:st=0:d=0.5:alpha=1,fade=t=out:st=40.5:d=0.5:alpha=1,setpts=PTS+40.5/TB[vpost-t];
[apost]atrim=start=795:duration=28,asetpts=PTS-STARTPTS,afade=t=in:st=0:d=0.5,afade=t=out:st=27.5:d=0.5,adelay=40500|40500[apost-t];
[v3]trim=start=95:duration=30,setpts=PTS-STARTPTS,format=yuva420p,fade=t=in:st=0:d=0.5,setpts=PTS+41+28-0.5/TB[v3-t];
[a3]atrim=start=95:duration=30,asetpts=PTS-STARTPTS,afade=t=in:st=0:d=0.5,adelay=68500|68500[a3-t];
[vpre-t][vpost-t]overlay[v1];
[v1][v3-t]overlay[v];
[apre-t][apost-t][a3-t]amix=inputs=3[a0];
[a0]loudnorm[a]" -map "[v]" -map "[a]" -y -c:v libx264 -preset ultrafast -b:v 15000k -aspect 1920:1080 -c:a aac fullgame_Output.mp4
Just do it in one command.
Besides the keyframe seek issue, which is true, your present sequence has an error in the last command. You have [0:v]trim=start=0.5...[0c] which trims out the first 0.5 seconds and will cause a desync of its own. Since this is the first clip, it should be [0:v]trim=0:40.5.
The full single command should be
ffmpeg -i fullgame.mp4 -filter_complex
"[0]split=2[vpre][vpost];[0]asplit=2[apre][apost];
[vpre]trim=start='00:09:57':duration='00:00:41',setpts=PTS-STARTPTS[vpre-t];
[apre]atrim=start='00:09:57':duration='00:00:41',asetpts=PTS-STARTPTS,afade=t=out:st=40.5:d=0.5[apre-t];
[vpost]trim=start='00:15:45':duration='00:00:28',setpts=PTS-STARTPTS,format=yuva420p,fade=t=in:st=0:d=0.5:alpha=1,setpts=PTS+40.5/TB[vpost-t];
[apost]atrim=start='00:15:45':duration='00:00:28',asetpts=PTS-STARTPTS,afade=t=in:st=0:d=0.5[apost-t];
[vpre-t][vpost-t]overlay[v];
[apre-t][apost-t]acrossfade=d=0.5,loudnorm,aresample=48000[a]"
-map "[v]" -map "[a]" -c:v libx264 -preset ultrafast -b:v 15000k -aspect 1920:1080 -c:a aac fullgame_Output.mp4
Your original sequence had -strict -2 for audio AAC encoding. That hasn't been needed since Dec 2015. You have a very old version of ffmpeg if your ffmpeg throws an error without it. Upgrade first.
I did not test the above with your file, as it will take too long to filter 16 min of Full HD 60 fps video, but I tested the below faster command and it works fine with the latest git build of ffmpeg:
ffmpeg -ss 00:09:57 -t 00:00:41 -i fullgame.mp4 -ss 00:15:45 -t 00:00:28 -i fullgame.mp4 -filter_complex
"[0]afade=t=out:st=40.5:d=0.5[apre-t];
[1]format=yuva420p,fade=t=in:st=0:d=0.5:alpha=1,setpts=PTS+40.5/TB[vpost-t];
[1]afade=t=in:st=0:d=0.5[apost-t];
[0][vpost-t]overlay[v];
[apre-t][apost-t]acrossfade=d=0.5,loudnorm,aresample=48000:ocl=stereo[a]"
-map "[v]" -map "[a]" -c:v libx264 -preset ultrafast -b:v 15000k -aspect 1920:1080 -c:a aac fullgame_Output.mp4
I use following encoding option.
ffmpeg -i input.wmv -movflags faststart -c:v libx264 -profile:v baseline -acodec aac -ac 2 -ar 48000 -strict -2 /root/output.mp4;
But sometimes last 5~30 seconds audio is erased.
Total video duration is about 3 mins.
what do you think the problem is?
Is this related to computer performance? I use quadcore, 4G ram.
I am trying to implement HLS using FFmpeg for transcoding + segmenting but have been facing a couple of issues that have been bugging me for the past week.
Issue
Webserver currently receives live MP4 fragments being recorded on-the-go and needs to take care of transcoding and segmentation.
As mp4 fragments are being received, they need to be encoded. Then segmented. If i run a segmenter (be it ffmpeg or apple mediastreamsegmenter), every mp4 fragment is being treated as a VOD by itself and I'm not being able to integrate them as part of a larger live event implementation.
I thought of a solution where every time I receive an mp4 fragment, I first use fmpeg to concatenate it with previous ones to form the larger mp4 that I then pass to be segmented for HLS. That did not work either because the entire stream has to be re-segmented each and every time and existing TS fragments replaced by new ones that are similar yet shifted in time.
Implementation 1
ffmpeg -re -i fragmentX.mp4 -b:v 118k -b:a 32k -vcodec copy -preset:v veryfast -acodec aac -strict -2 -ac 2 -f mpegts -y fragmentX.ts
I manage the m3u8 manifest on my own, deleting old fragments and appending new ones.
When validating the stream, I find it stacked with EXT-X-DISCONTINUITY tags making the stream unwatchable.
Implementation 2
First combine latest fragment with overall.mp4
ffmpeg -i "concat:newfragment.mp4|existing.mp4" -c copy overall.mp4
Then pass the combination to ffmpeg for HLS segmentation
ffmpeg -re -i overall.mp4 -ac 2 -r 20 -vcodec libx264 -b:v 318k -preset:v veryfast -acodec aac -strict -2 -b:a 32k -hls_time 2 -hls_list_size 3 -hls_allow_cache 0 -hls_base_url /Users/JosephKalash/Desktop/test/350/ -hls_segment_filename '350/fragment%03d.ts' -hls_flags delete_segments 350/index.m3u8
Concatenation is not perfect and there are noticeable glitches where the fragments are supposed to be stitched. Segmentation replaces older fragments and the manifest is rewritten as if it's a new HLS stream every time ffmpeg is called.
I cannot figure out how to get this to work properly.
Any ideas?
Solved by relying on nginx rtmp module which turned out to be suited for the above implementation.
I am using the following command to encode an AVI to an H264 video for use in an HTML5 video tag:
ffmpeg -y -i "test.avi" -vcodec libx264 -vpre slow -vpre baseline -g 30 "out.mp4"
And this works just fine. But I also want to create a placeholder video (long story) from a single still image, so I do this:
ffmpeg -y -i "test.jpg" -vcodec libx264 -vpre slow -vpre baseline -g 30 "out.mp4"
And this doesn't work. What gives?
EDIT: After trying LordNeckbeards answer, here is my full output: http://pastebin.com/axhKpkLx
Example for a 10 second output:
ffmpeg -loop 1 -framerate 24 -i input.jpg -c:v libx264 -preset slow -tune stillimage -crf 24 -vf format=yuv420p -t 10 -movflags +faststart output.mp4
Same thing but with audio. The output duration will match the input audio duration:
ffmpeg -loop 1 -framerate 24 -i input.jpg -i audio.mp3 -c:v libx264 -preset slow -tune stillimage -crf 24 -vf format=yuv420p -c:a aac -shortest -movflags +faststart output.mp4
-loop 1 loops the image input.
-framerate sets the frame rate of the image input. Default is 25. Some players have issues with low frame rates so a value over 6 or so is recommended.
-i input.jpg the input.
-c:v libx264 the H.264 video encoder.
-preset x264 encoding preset. Use the slowest one you can.
-tune x264 tuning for various adjustments to fit specific situations.
-crf for quality. A lower value results in higher quality. Use the highest value that still provides an acceptable quality to you. Default is 23.
-vf format=yuv420p outputs the pixel format as yuv420p. This ensures the output uses a widely acceptable chroma sub-sampling scheme. Recommended for libx264 when encoding from images.
-c:a aac the AAC audio encoder. If your input is already AAC or M4A then use -c:a copy instead to stream copy instead of re-encode.
-t 10 (in the first example) makes a 10 second output. Needed because the image is looping indefinitely.
-shortest (in the second example) makes the output the same duration as the shortest input. In this case it is the audio since the image is looping indefinitely.
-movflags +faststart relocates the moov atom to the beginning of the file after encoding is finished. Allows playback to begin faster in progressive download playing; otherwise the whole video must be downloaded before playing.
-profile:v main (optional) some devices can't handle High profile.
See FFmpeg Wiki: H.264 for more info.