Make batched suppress_tokens behaviour same as in sequential

[Purfview]

Make batched suppress_tokens behaviour same as in sequential mode.

[MahmoudAshraf97]

They should have the same behavior, the else case is already handled in the function itself

faster-whisper/faster_whisper/transcribe.py

Lines 1836 to 1837 in 8327d8c

	elif suppress_tokens is None or len(suppress_tokens) == 0:
	suppress_tokens = [] # interpret empty string as an empty list

I'd suggest modifying the sequential code instead for consistency

[Purfview]

They should have the same behavior, the else case is already handled in the function itself

Not same, as the function populates suppress_tokens.
I think original author had in mind an option to have suppress_tokens empty.

Look at #898

[MahmoudAshraf97]

I see, but I think the batched behavior is still the correct one, as you can see here, there are a certain set of tokens that are always suppressed regardless of the user request, these tokens are task tokens or prompt tokens so we should not expect them to be generated in any case

[MahmoudAshraf97]

@jordimas

[jordimas]

Hello!
I have very little time to check the code, but from what I remember there are 3 uses cases:

• User provides None tokens (nothing is applied)
• User provides -1 and then the defaults are used
• The user provides a list and then the defaults + user's list is used

Consider also that adding suppress is not called is the if evaluates to False (e.g. None)
https://github.com/openai/whisper/blob/90db0de1896c23cbfaf0c58bc2d30665f709f170/whisper/decoding.py#L557C25-L557C40

What I can tell is that I test extensively OpenAI and faster-whisper to behave the same in all cases when I did the PR.
I think that @Purfview is right.

[MahmoudAshraf97]

Thanks all, I will merge then