AI assisted transcription and subtitle creation

[lsg1310]

For whatever reason, I've had little to no luck with using Whisper through SubtitleEdit to generate subtitles, but so far, this site seems to work really well and has some nice options.

https://turboscribe.ai/

[DreamScape]

For whatever reason, I've had little to no luck with using Whisper through SubtitleEdit to generate subtitles

SubtitleEdit has like 10 different engines these days. None of them work? Do they crash or are they not detecting any speech?

[Night457]

That site promises transcription in SECONDS? The last time I attempted SubtitleEdit Whisper, it was on a short film maybe 15 minutes long, it took about 8 hours, and the results were idiotic so I scrapped them.

My suspicions are simply about the speed, not about someone other than me getting good results.

[lsg1310]

Pretty much every time I tried it in Subtitle Edit, it took forever and I would get repeated lines. Maybe I just wasn't picking the right option or something, but this seems to work a lot better for me. If you wanted to try it for free, you get three transcriptions per day and I think the limit is 30 minutes

[DreamScape]

Unfortunately repeating text is a common problem for all of us. The solution is usually to spend more time by trying a different model or cutting the film into smaller pieces that you'd then recombine when finished.

[lsg1310]

Yeah, I just never tried tweaking with all the options

[Night457]

cutting the film into smaller pieces that you'd then recombine when finished.

That sounds like a technique that could address a number of issues, thanks. But I have a dumb question: do the various AI transcription methods work better with loading audio only and not video? Or does that only affect how long it takes to (up)LOAD the file, because just audio would be a smaller file? I ask because SubtitleEdit is geared toward full video files, not specifically audio. But I can see how this would definitely be an issue with online services, because uploading a 5Gb video would definitely take longer than a 250Mb audio file.

Incidentally, TurboScribe says this:

10 Hour Uploads
Each file can be up to 10 hours long / 5 GB. Upload 50 files at a time.

They just might be a little overly confident in speed and accuracy if your experience with a 90-minute movie requires cutting it into pieces and THEY claim 50 files at 5Gb and 10 hours long each is just fine!

(...does anyone think the TurboScribe site text is AI-generated?)

[DreamScape]

But I have a dumb question: do the various AI transcription methods work better with loading audio only and not video? Or does that only affect how long it takes to (up)LOAD the file, because just audio would be a smaller file? I ask because SubtitleEdit is geared toward full video files, not specifically audio. But I can see how this would definitely be an issue with online services, because uploading a 5Gb video would definitely take longer than a 250Mb audio file.

Whisper only processes audio, so there shouldn't be a difference in the final output.

[ghost]

I figured out, that Whisper in SubtitleEdit works best with Purviews's Faster-Whisper. I gives you no repetitions and is the fastest engine.

[Image]

Pretty much every time I tried it in Subtitle Edit, it took forever

You should have really strong hardware and the main thing: a fast graphic card like a Nvidia RTX card to get good speeds with Whipser.

[Night457]

Whisper in SubtitleEdit works best with Purviews's Faster-Whisper.

Well, at least I got that part right.