Does faster-whisper-xxl alter the use of the best_of parameter?

[subgrinder]

from faster-whisper-xxl --help:

--best_of BEST_OF, -bo BEST_OF
number of candidates when sampling with non-zero
temperature (default: 5)

---

from Whisper API Docs (https://whisper-api.com/docs/transcription-options/):

Best of

This parameter determines how many of the highest-scoring beam search results to consider for the final output. It works in conjunction with beam_size to balance between exploration and final selection quality.

The default value is 5.

[Purfview]

Does faster-whisper-xxl alter the use of the best_of parameter?

No idea what your question means, but best_of scoring code is different (theoretically better) than vanilla whisper's, anyway, best_of is random thing. So, probably impossible to measure what is better.

BTW, as you do benchmarks/comparisons then you don't want this parameter (randomness) to be in use at all. Use -fallback None!

[subgrinder]

According to the Whisper API Docs, the "best_of" parameter is used with beam search.

According to faster-whisper-xxl:
"best_of" is used with temperature when settings are non-zero. Beam size and temperature non-zero are mutually exclusive - only beam search or temperature non-zero can be used in a transciption.

So, did faster-whisper-xxl change the use of best_of to work with temperature instead of beam size?

[Purfview]

According to the Whisper API Docs, the "best_of" parameter is used with beam search.

It's poorly worded or outdated. They probably meant "used together" in a general sense, but in practice the two parameters don’t interact directly because they control two different decoding modes. Model can restart decoding segments with "best_of" mode if a segment fails or produces low-quality text.

BTW, another thing I noticed in those docs: "The default beam size is 5". As I remember it's true only to cli, but when used directly in Python then by default it's using the greedy search (aka beam_size=1).

So, did faster-whisper-xxl change the use of best_of to work with temperature instead of beam size?

It works same as in vanilla whisper, just the "best_of" scoring method is different.