Whisper cuts timestamp every seconds

[701789262a]

Using this code results in having every timestamp trimmed in seconds (the segment start and ends ONLY with integers)

model = whisper.load_model("large-v2")

options = dict(without_timestamps=False)
transcribe_opt = dict(task="transcribe",**options)
result = model.transcribe("audio.wav",**transcribe_opt)

I've seen that it's possible to use stable-ts but I was wondering if was missing something (as if stable-ts only optimizes it but whisper itself should be able to return decimal timestamp)

[jongwook]

The model has a bias on integer timestamps. Specifying word_timestamps=True will recalculate the timestamps based on the predicted word-level timestamps, and it tends to be more accurate.

Related:

• Improving Timestamp Accuracy #435
• The time precision in generated subtitle files is second, not millisecond #809
• Verbose Response should include exact (not integer rounded) timestamps #1122