setting initial_prompt seems to override suppress_tokens or 'unsuppress' speech disfluencies (not sure which)

[filmo]

When I set initial_prompt='Some phrase here' I end up with ums and uhs throughout the transcription.

Adding the token_ids for 'um', 'Um', 'uh', and 'Uh' to suppress_tokens= [449, 3232, 40937, 27727] doesn't eliminate them.

I tried adding these same token_id to the 'suppress_ids' set in config.json and leaving suppress_tokens set to default (-1) and that doesn't make a difference either.

Is there something about using initial_prompt that is causing either suppress_tokens to be ignored, or promoting the model to include the speech disfluencies?? I'm not sure which is it is, but no matter what prompt I use, I end up with lot of ums and uhs in the resulting transcription.

Thank you in advance for any thoughts or guidance you might have on this.