Infinite loop to previous segment timestamp

[dodysw3]

Hi, sometimes whisper transcription stuck in an infinite loop where it went to past segment timestamp and never get out of it. E.g.

...
[38:40.140 --> 38:49.740]  Terus ada beberapa fitur juga di dalamnya yang sangat-sangat support sekali untuk di tahun 2022 ini.
[38:49.740 --> 38:34.860]  Di mana, Desk
[38:34.860 --> 38:40.140]  Tapi untuk resolusi 2K pun itu udah bagus, cuma ada yang lebih bagus lagi itu 4K.
[38:40.140 --> 38:49.740]  Terus ada beberapa fitur juga di dalamnya yang sangat-sangat support sekali untuk di tahun 2022 ini.
[38:49.740 --> 38:34.860]  Di mana, Desk
...

whisper parameters:
whisper input.wav --language id --model medium --condition_on_previous_text False --no_speech_threshold 0.4 --best_of None --beam_size None --device cuda:0

Any idea?

[junkoran]

solved blm bang?

[andrewchernyh]

I've catched something simular, and as I see it's related to

whisper/whisper/transcribe.py

Line 216 in 5c1a8c1

last_timestamp_position = (

When decoder has misprediction of the latest consecutive timestamp, in my case it was tokenizer.timestamp_begin seek changes is wrong and transcribe can't be finished.

I'm not sure, but I think decoder should suppress emitting of timestamp tokens with ts <= latest timestamp.
Something like

whisper/whisper/decoding.py

Line 430 in 5c1a8c1

if self.max_initial_timestamp_index is not None:

, but according to the current state

[andrewchernyh]

Fix in #914

[doublex]

Still fails: #924

[Jeronymous]

This problem is still here even after the fix by @andrewchernyh
I report precisely with @doublex in #924