There will be a phenomenon that the subtitles are identified and repeated about whisper

[Const-me]

@1353604736 Windows is usually fine sharing hardware resources between multiple running processes. And I think when the issue is interference with other running processes, the symptoms should be different. For CPU load, it’s just slow down. For GPU load one possible symptom is TDR timeout, like DXGI_ERROR_DEVICE_REMOVED code from some GPU API method.

Might be a bug somewhere.
Can you upload or link the audio you’re processing?

[1353604736]

@1353604736 Windows is usually fine sharing hardware resources between multiple running processes. And I think when the issue is interference with other running processes, the symptoms should be different. For CPU load, it’s just slow down. For GPU load one possible symptom is TDR timeout, like DXGI_ERROR_DEVICE_REMOVED code from some GPU API method.

Might be a bug somewhere. Can you upload or link the audio you’re processing?

Similar issue found: ggerganov/whisper.cpp#471
许老脉诊12_MP3.zip

[1353604736]

mp3 file to reproduce the problem
张震号脉01_0002_MP3.zip

[albino1]

@Const-me, looking through whisper.cpp issues it seems like a lot of people solved this repeated/duplicated lines problem by adjusting the entropy threshold. For example:

ggerganov/whisper.cpp#471 (comment)
ggerganov/whisper.cpp#508 (comment)

I'm wondering if there would maybe be a way to give access to the --entropy-thold value within your GUI.

For me, even using version 1.7, I still get endless repetitions in almost anything I try that's not English. If I split it into small enough chunks I can get it to work, but eventually it breaks down. Sometimes it recovers, and starts working, but then inevitably it will start duplicating lines again.

Here's an Italian test file example with the model set to medium.bin and the language set to Italian where at 2:24 it just starts saying Grazie. over and over:

italian-test.zip

test-medium.txt

Switching the model to large.bin results in a similar behavior, except this time starting at 2:33 it just starts repeating [Musica] over and over and never recovers:

test-large.txt

Using tiny gives a similar problem, but it eventually starts repeating Ma la mia guarda for the final 2+ minutes:

test-tiny.txt

I originally tested this with a WAV file, but switched to an MP3 and redid all my tests so the filesize would be small enough to post to Github.

[DhruvDh]

I have been experiencing a lot more of these issues on the newer version of WhisperDesktop. Lot's of "[BLANK], [INAUDIBLE], [COUGH]", and lots and lots of Okays repeating. Reverting to a release I downloaded on 02-14 solves this.

[DhruvDh]

I see this issue again even after downgrading - maybe something went wrong when I updated my drivers. 7900 XTX on adrenalin 23.3.1. At some point I will try it in Linux with Wine and report back.