Hullucinating issues

[dubefab]

When I try transcribing (especially with large-v3) an English .wav that's about 60 minutes long. If there's a song somewhere in the video that repeats the same thing, it will start getting stuck in a loop of transcribing the same thing over and over again and won't even finish the transcription. Is there something I can do about this? Like options to enable? Thanks

[marekves]

I have the same issue, could you solve it?

[taivlam]

I've also noticed that large-v3* overall performs worse than large-v2*.

Either large-v3* works until it experiences a repetitive hallucination at some point into the audio file (some examples: 3 min into a 2 hr recording, and 40 min into a 1 hr recording); or large-v3* avoids a repetitive hallucination loop at the cost of performing overall worse than even large-v1*.

See the following for reference:

• GH issue Whisper large v3 model repeats a lot #1507 from this repo
• Discussion from openai/whisper#1762 (comment)
• GH discussion Repeating text after ~40 minutes #1911 from this repo