Please help! Getting only one character as output.

[Pelad0x0]

Hello! I'm trying to use the cli and I'm getting a wrong output.
I'm using Proxmox and a Lxc container (Debian 12) with a GPU (Gforce 3060ti) passthrough. Then compiled python 3.11 and installed whisper with /opt/py311/bin/python3 -m pip install git+https://github.com/openai/whisper.git.

When using /home/sysoxo/.local/bin/whisper --model small --language es --word_timestamps True --output_format srt Trim.mp4
the output is correct:

The issue comes when I try to use the medium model /home/sysoxo/.local/bin/whisper --model medium --language es --word_timestamps True --output_format srt Trim.mp4 I'm getting this output:

nvidia-smi output:

What I'm missing? Thank you very much.
Martin.

[glangford]

Note the difference between what is displayed on the console (which is just a log) compared to what is output in the actual .srt file. Also, try doing the same thing with the medium model without --word_timestamps.

[Pelad0x0]

I've tried without --word_timestamps with the same result. And the content of .srt file is the same as the output. Only i.
Thanks for the response.

[glangford]

That's unfortunate, what you are seeing is probably hallucination that causes a 'stutter' in the transcript. See for example these discussions

• A possible solution to Whisper hallucination #679
• How to prevent hallucinations/time-shifts? #919

[Pelad0x0]

Thank you glangford! Those links lead me to the solution. I've added --logprob_threshold None to the command.

whisper --task transcribe --model medium --language es --output_format srt --word_timestamps True --logprob_threshold None Trim.mp4