Problem with time stamp and audio size

[titusfx]

I've an audio that is 9 seconds long, whisper returns a transcription with 34 seconds long.

The audio has 4 phrases.
The audio is mp3 file and the result is on the txt file (json file)

audio_example
whisper_result

[jongwook]

We could add a post-processing step so that the timestamps are upper-bounded by the audio length. This wasn't my priority because the timestamps tend to become more accurate with the larger models and a prolonged last timestamp didn't hurt much practically.

[kmk2018]

Hello,
This problem can occur within an utterance if Whisper breaks the utterance into multiple phrases, not just the final end time. For example this is from a 11.09s file with 4 phrases where the 3rd and 4th phrase have the wrong timings:

["cand", [{"id": 0, "seek": 0, "start": 0.0, "end": 4.6000000000000005, "text": ...}, {"id": 1, "seek": 0, "start": 4.6000000000000005, "end": 8.64, "text": ...}, {"id": 2, "seek": 864, "start": 8.64, "end": 14.66, "text": "...}, {"id": 3, "seek": 864, "start": 14.66, "end": 34.36, "text": ...}]],

FYI the input was a .wav file

[Pricer187]

We could add a post-processing step so that the timestamps are upper-bounded by the audio length. This wasn't my priority because the timestamps tend to become more accurate with the larger models and a prolonged last timestamp didn't hurt much practically.