Open hinswhale opened 1 month ago
I just started seeing this. Did you per chance recently start using a different whisper model?
There is an opened discussion on this : https://github.com/linto-ai/whisper-timestamped/discussions/79#discussioncomment-10405887
It seems to be a corner case, that happens when the Whisper model predicts a transcript which only involves special language tokens up to the maximum token length (e.g. <|0.00|><|de|><|de|><|de|><|de|><|de|>...
).
I am just waiting to have a quick way to reproduce this corner case, to be able to fix it safely.
I met this problem several times,what can I do to fix it? Thanks Perhaps we should implement a feature to temporarily save transcribed files, allowing us to double-check the results and ensure that previous work isn't lost