Using Purview Faster Whisper in Subtitle Edit

[GrampaWildWilly]

As instructed here niksedk/subtitleedit-avalonia#13 (comment) . . .

During install of PFW in the Subtitle Edit Speech to text dialog, I saw a new ffmpeg being downloaded. I already had ffmpeg on my system. Is there some reason you don't just use the one I already have? Do you have your own local mods? What harm would there be in my copying my ffmpeg into the PFW directory, replacing the one that's already there?

If you update PFW, how does SE learn of it & get the new version? Ditto for any models.

I was misled by the PFW documentation into thinking --beep_off true was the correct syntax. Alternative wording could be that you either supply or omit --beep_off, default being omitted, & the parameter takes no parameter value.

At the completion of a transcription in SE, PFW saves the generated subtitle file in the PFW directory. I suggest it be placed instead in a subtitle directory within the PFW directory. I also suggest that there be some management tool that does either of these. (1) When the user saves the generated subtitle file, PFW (or SE?) deletes the file PFW saved. (2) Failing that, there needs to be some way of aging the PFW-saved subtitles & deleting them at some expiration date.

The whisper_log.txt file just grows & grows with each transcription in SE. There needs to be a tool that pares the file automatically on some sort of schedule so it doesn't just grow to infinity. Maybe if it gets to be over 50,000 lines (or some other threshold), you automatically delete the last 45,000 of them before you start logging the current activity into the front of the file.

I used --hallucination_silence_threshold 4 & it suppressed lots of hallucinatory subtitles. Thanks for that.

[Purfview]

During install of PFW in the Subtitle Edit Speech to text dialog, I saw a new ffmpeg being downloaded. I already had ffmpeg on my system. Is there some reason you don't just use the one I already have? Do you have your own local mods? What harm would there be in my copying my ffmpeg into the PFW directory, replacing the one that's already there?

For ffmpeg.exe, it looks in the same folder, or in the PATH, or in the system folder. Where and what version you want to use - is up to you.

If you update PFW, how does SE learn of it & get the new version? Ditto for any models.

nikse updates SE with new links/hashes. Models don't get updated, they come with new names like large-v2/large-v3 ect.

I was misled by the PFW documentation into thinking --beep_off true was the correct syntax.

--beep_off is a switch, it doesn't take any arguments. Yes, Python's default help is not very intuitive at the first glance.

At the completion of a transcription in SE, PFW saves the generated subtitle file in the PFW directory.

PFW already does that by default. How SE does it - it's up to SE.

The whisper_log.txt file just grows & grows with each transcription in SE.

This is SE related.

I used --hallucination_silence_threshold 4 & it suppressed lots of hallucinatory subtitles. Thanks for that.

Yes, it can help, it's similar to the original Whisper functionality [I tweaked the logic a bit to reduce false positives and added some additional functionality].
You can use shorter command - -hst, I would recommend "2" value for it.

[GrampaWildWilly]

Thanks. Nik's got my discussion in his repository. I like to use long names instead of abbreviations. I will surely forget what the abbreviation is an abbreviation of.

[Purfview]

BTW, ffmpeg is not in use by default, it's used only with --ff_... filters.

[GrampaWildWilly]

Aha. In that case the issue is moot, at least for me. I don't do fancy stuff with it. I'll probably never code those parameters so I'm not going to do anything about ffmpeg in PFW. Thanks again.