Transcript lines get shorter and shorter until one word per line

[ecotechie]

I'm on version 20240930 and am running this command:
whisper --output_format all --model turbo --language English --threads 10 file.mp4

Everything seems fine at the beginning. Each line is about a sentence long. But then, slowly, the lines start getting shorter (not related to sentence length). Eventually, half way through or so, all lines are one word. I've tested this several times on different files.

Has anybody seen this, or know how to "force" line length?

[ryanheise]

Whisper has a tendency to continue transcribing in the style of what came prior. Most of the time this works well, but it can sometimes cause Whisper to get stuck in an undesirable pattern. I suppose you could try the --condition_on_previous_text False option to prevent this happening, but then you would also lose the benefits of Whisper's context awareness.

On the other hand, in the case of line length, yes there is a way to "force" line length without modifying the previous option. If you go to the README page and read the section on "Command-line usage", there is a part that says:

Run the following to view all available options:

whisper --help

Typing this, you can see all the options related to line length.

In short, try:

your_command --word_timestamps True --max_line_width 44 --max_line_count 1

[eroo36]

I tried --word_timestamps True --max_line_width 44 --max_line_count 1 on the terminal, it was still doing one word per line after some point, but I looked at the generated .srt file, and it looked okay there, it was not like that.

[ryanheise]

That is the correct behaviour. When you see the transcription progress printed on the screen while it's transcribing, that is not intended as final output, that is just the progress indicator. The final transcript output gets written to the SRT file (and/or VTT file depending on which output format option is specified).