Patience value in beam decoding

[glangford]

There is an optional patience parameter from the command line:

  --patience PATIENCE   optional patience value to use in beam decoding, as in https://arxiv.org/abs/2204.05424, the default (0.0) is equivalent to not using patience (default: 0.0)

In the referenced arxiv paper, "Beam Decoding with Controlled Patience", the patience factors referenced there will not translate directly (pun intended) to whisper. This is because whisper sets its patience as the command line value + 1, at least according to

whisper/whisper/decoding.py

Line 283 in 520796a

self.max_candidates: int = round(beam_size * (1.0 + patience))

I just wanted to point out the difference for anyone else experimenting with the option.

[jongwook]

Thanks for pointing this out! This must have been my overlook, and I should change the default / conventional to 1.0. This would be a breaking change but hopefully no one minds.