Fact · Transcription
In Whisper's transcribe function, a segment counts as silent when its no-speech probability is above the no_speech_threshold setting and its average log probability is below logprob_threshold.
Checked against its source on October 17, 2026
In the source's own words
“If the no_speech probability is higher than this value AND the average log probability over sampled tokens is below ... consider the segment as silent”
whisper/transcribe.py: decoding defaults and command-line options (openai/whisper), OpenAI, GitHub (openai/whisper), 2026, transcribe() parameter documentation in whisper/transcribe.py.
Open the original- What it measures
- The documentation sits beside the decoding options that control how Whisper retries a segment it is unsure of. The threshold is one of the settings a user can change when quiet passages produce invented text.
- What it does not show
- The setting is a tuning control, not a guarantee. Changing it can drop real speech as well as invented text, so check the result against the audio.