Fact · Speech recognition
The openai-whisper command line turns its starting temperature and its fallback increment into a list of temperatures that runs up to 1.0, which the decoder tries in turn when a decode fails.
Checked against its source on October 17, 2026
In the source's own words
“temperature = tuple(np.arange(temperature, 1.0 + 1e-6, increment))”
whisper/transcribe.py: decoding defaults and command-line options (openai/whisper), OpenAI, GitHub (openai/whisper), 2026, Command-line options, after argument parsing.
Open the original- What it measures
- The increment comes from the --temperature_increment_on_fallback option. Setting that option to nothing leaves a single temperature, so the fallback is turned off.
- What it does not show
- This is a line of source code, not a published benchmark. It shows how the list is built, not how much each retry improves the transcript.