Problem with repetitions

[lukasc-ubc]

I am transcribing audio to text using Whisper and the Aiko app. About 30% of the way through a 1 hour recording, it gets stuck in a loop, and outputs:
Thank you. Thank you. Thank you.

I tried a setting in the app, "Reduce repetitions", and it get a different repetitive pattern, again stuck in a loop.

The above was on Mac OSX M1 processor, using v2 of the model.

I also tried with the iPhone. and it got stuck on "Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, Uh, "

[glangford]

Unfortunately this is a well known issue that hasn't been resolved. See this past discussion for an explanation and possible workarounds.

• Some wired repetition happens in transcription. #192

[lukasc-ubc]

@glangford

Thank you for your reply. That discussion topic was 2 years ago, and doesn't seem to have a resolution. I had a colleague run my audio file using the Python API, and it also had problems, but generated some different repetition text. Shouldn't this have been fixed in the model?

I imagine the Aiko app uses that suggestion in their option for reducing repetitions, but I'll ask.

[glangford]

Shouldn't this have been fixed in the model?

There are new models you can experiment with (large-v3, turbo) but the underlying problem has not been fixed as far as I am aware.

[Purfview]

Check this bugfix: #1903