without_timestamps significantly changes the result (for the worse).

[jzju]

There is 50% more words for without_timestamps=0 compared to without_timestamps=1.

[phineas-pta]

consider adding a reproducible example in a bug report to help devs figuring out what happen

[jzju]

model = WhisperModel("medium", device="cuda", compute_type="float16")
f = "o.mp3"
segments, info = model.transcribe(f, beam_size=1, without_timestamps=1, language="en")
print(" ".join([s.text.strip() for s in segments]))
print()
segments, info = model.transcribe(f, beam_size=1, language="en")
print(" ".join([s.text.strip() for s in segments]))

For this clip there is a small differnce https://soundbible.com/1360-Obama-State-Of-The-Union-2010.html
I much more difference for a 20min clip in swedish with multiple speakers and noice but it's not data I could publish.

tear each other down instead of lifting the bar, lifting this country up, we lose faith.
vs
tear each other down instead of lifting this country up, we lose faith.

[phineas-pta]

more likely hallucinations with the speech above

try use large model and/or VAD

[guillaumekln]

You could also try with the original openai-whisper but it is generally expected that without_timestamps changes the results. The decoding logic is very different.