Fact · Speech recognition
Whisper's transcribe code treats a segment as silence, and skips it, when its no-speech probability is above one threshold and its average log probability is below another.
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 `logprob_threshold`, consider the segment as silent”
whisper/transcribe.py: decoding defaults and command-line options (openai/whisper), OpenAI, GitHub (openai/whisper), 2026, Docstring for the no_speech_threshold argument of the transcribe function, in whisper/transcribe.py.
Open the original- What it measures
- The function's defaults are no_speech_threshold 0.6 and logprob_threshold -1.0. The skip check runs on each 30-second window the model decodes, so a stretch of quiet audio can produce no text at all.
- What it does not show
- This is the reference implementation's rule. Other tools built on Whisper can set these thresholds differently, and the rule cannot reliably separate very quiet speech from silence, so it can drop quiet speech too.