Non-increasing timestamps bug?

[tobyshooters]

Near the end of a local file, I'm getting these peculiar results.
Still looking through the decoding logic for timestamp inference, but though worth noting here:

               [04:39.960 --> 04:40.960]  This is a sample.
               [04:40.960 --> 04:41.960]  Ciao, ciao.
               [04:41.960 --> 04:43.960]  What's the best fashion I've ever received?
WRONG TS:      [04:43.960 --> 04:18.000]  You wear the
HALLUCINATION: [04:18.000 --> 04:19.000]  crown.
               [04:19.000 --> 04:20.000]  What's the most memorable thing that you've ever worn?
               [04:20.000 --> 04:22.000]  Met Gala Award, Taxitio from Gucci with a diamond and
               [04:22.000 --> 04:24.000]  crossed it's sort of a cobra on the back.
               [04:24.000 --> 04:25.000]  That was pretty sick.
               [04:25.000 --> 04:26.000]  What do you do on your off days?
               [04:26.000 --> 04:27.000]  Nothing.
               [04:27.000 --> 04:28.000]  Just take it easy.
               [04:28.000 --> 04:29.000]  Run around with the kids.
               [04:29.000 --> 04:30.000]  Hey.
               [04:30.000 --> 04:32.000]  I'm just more than a man.
               [04:32.000 --> 04:33.000]  This is a sample.
               [04:33.000 --> 04:36.000]  Ciao ciao.
               [04:36.000 --> 04:37.000]  What's the best fashion I've ever received?
               [04:37.000 --> 04:38.000]  You wear the crown.
               [04:38.000 --> 04:39.000]  Ciao.
               [04:39.000 --> 04:41.000]  What's the best fashion I've ever received?
               [04:41.000 --> 04:42.000]  You wear the crown.
BACK ON TRACK: [04:42.000 --> 04:44.000]  What's the best fashion I've ever received?
               [04:44.000 --> 04:45.000]  You wear the crown.
               [04:45.000 --> 04:47.000]  I see you've ever received.
               [04:47.000 --> 04:51.000]  You wear the clothes, not the clothes, where are you?

Repro:

git clone https://github.com/openai/whisper
cd whisper
*copy file into directory*
ipython
>> import whisper
>> model = whisper.load_model("tiny.en", device="cpu")
>> result = whisper.transcribe(model, "federer_short.mp4", language="en", verbose=True)

Audio file federer_short.mp4: https://we.tl/t-GkGifgJCuP

[Joshfindit]

Similar stuff here, using an mp3 of a one-side-remote 1:1 meeting.

[jianfch]

you can try to use beam search instead of the default greedy search which more prone to this failure mode