Fact · Speech recognition
25milliseconds per window
Whisper re-samples all audio to 16,000 Hz and computes its 80-channel log-magnitude Mel spectrogram on 25-millisecond windows with a stride of 10 milliseconds.
Checked against its source on October 17, 2026
In the source's own words
“All audio is re-sampled to 16,000 Hz... is computed on 25-millisecond windows with a stride of 10 milliseconds”
Robust Speech Recognition via Large-Scale Weak Supervision (full text), Alec Radford et al., OpenAI (arXiv), 2022, Section 2, Approach (data processing).
Open the original (DOI 10.48550/arXiv.2212.04356)- What it measures
- The Whisper paper describes the input its encoder receives for every audio clip, in its model description.
- What it does not show
- This is the front end of the Whisper models in the paper. Other engines use different windows, and the numbers describe the input, not the accuracy.