fix: use default whisper temperature values for temperature 0.0 like openai api

[dotmobo]

Hi,

It's an alternative version of #553 made for the latest v0.9.0 RC.

So, like the OpenAI API, the default temperature remains 0.0. But OpenAI made some temperature changes behind this, cf: https://developers.openai.com/api/reference/resources/audio/subresources/transcriptions/methods/create

If set to 0, the model will use log probability (https://en.wikipedia.org/wiki/Log_probability) to automatically increase the temperature until certain thresholds are hit.

Then we can see that the default Whisper temperature values are a list of temperatures, cf: https://whisper-api.com/docs/transcription-options/#sampling-temperature

The default temperature values are [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]

So, I think that OpenAI transforms the temperature 0.0 to [0.0, 0.2, 0.4, 0.6, 0.8, 1.0] to activate that "automatically increase the temperature until certain thresholds are hit" feature.

Here we are doing the same. So it will fix the bug in #553.

Regards,

Morgan

[Benju1]

+1 for this fix. We're running into the exact same issue with a German CT2 model (TheTobyB/whisper-large-v3-turbo-german-ct2).

Symptoms: A 1:21 minute audio recording is only transcribed up to ~30 seconds. The remaining segments are silently dropped.

Debug logs show:

Processing segment at 00:00.000           → OK
Processing segment at 00:30.000           → Log probability threshold is not met with temperature 0.0 (-1.614954 < -1.000000)
Processing segment at 01:00.000           → Log probability threshold is not met with temperature 0.0 (-1.994506 < -1.000000)

Because temperature is passed as a single float 0.0 (no fallback list), faster-whisper cannot retry with higher temperatures. Combined with no_speech_prob > no_speech_threshold, segments 2 and 3 are classified as silence and their text is completely dropped.

Workaround: We're currently using a sed-based entrypoint patch to set no_speech_threshold=None, log_prob_threshold=None, compression_ratio_threshold=None in the transcribe() calls, but the proper fix is this PR — converting temperature=0.0 to the default fallback list [0, 0.2, 0.4, 0.6, 0.8, 1.0] as OpenAI does.

This is a significant issue for non-English models where confidence scores are naturally lower. Would love to see this merged!

[Yichen-fqyd]

Thanks for submitting this pr
Would this replicate the same behavior of faster_whisper when compression_ratio hit compression_ratio_threshold, it will try different temperatures to help with hallucination.