Fact · Speech recognition
Whisper can retry a segment with a sequence of higher sampling temperatures when a decode fails its quality checks.
Checked against its source on October 17, 2026
In the source's own words
“which will be successively used upon failures according to either `compression_ratio_threshold` or `logprob_threshold`.”
whisper/transcribe.py: decoding defaults and command-line options (openai/whisper), OpenAI, GitHub (openai/whisper), 2026, transcribe() parameters: temperature.
Open the original- What it measures
- The default temperatures are 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. The command line defaults to temperature 0 and raises the temperature by 0.2 on each fallback.
- What it does not show
- Retries are a fallback for bad output, and they make a run slower. Setting a single temperature turns the fallback off.