Closed nhan000 closed 1 year ago
Check if .srt subtitle file is created. [when you think that it's "stopped"]
The srt file was created and the later half was missing, same as the timestamp in the command prompt.
Do you run it on cuda? If yes then try --compute_type=int8
parameter.
I added the parameter you gave me, so the code is
whisper-faster "C:\Users\ntnha\Videos\4K Video Downloader\Carl Sagan Astronomer of the People.mp4" --language en --model large-v2 --batch_recursive true --compute_type=int8
It ran on cuda
And it still stopped at the same location
I reproduced this issue on my side. Later I'll check what can be done about it. Interestingly, this hallucination starts on the advertisement.
Thanks for looking into this, and separately, thanks for making this program. Very noob-friendly for people who are not very techy like me.
The video has 3 advertisement segments:
It doesn't stuck with -beam_size=5
option.
Ad at start/end is still ignored, probably models are trained to ignore that ad. Btw tiny and base models transcribe that ad.
Thanks a lot! I will keep the beam size parameter in mind and change it around when I ran into issues.
I downloaded this 27 min Youtube video (uploaded it here).
I run the transcription using this code
whisper-faster "C:\Users\ntnha\Videos\4K Video Downloader\Carl Sagan Astronomer of the People.mp4" --language en --model large-v2 --batch_recursive true
and it stopped at
[13:15.860 --> 13:18.860] His greatest achievement was just around the corner.
I downloaded the mp3 file from that YouTube video (uploaded it here)
whisper-faster "C:\Users\ntnha\Videos\4K Video Downloader\Carl Sagan Astronomer of the People.mp3" --language en --model large-v2 --batch_recursive true
and it was able to run to
[26:44.760 --> 26:46.180] might have been enough.
Interestingly, it didn't transcribe the advertisement at the beginning and at the end of the video.