How it works
From sound to words
The audio is decoded in your browser, mixed down to one channel and resampled to 16 kHz, which is the rate Whisper listens at. Whisper, a speech model released by OpenAI under the MIT licence, then reads it in 30-second windows that overlap by five seconds and writes the words with a start and end time for each.
The model runs in a background thread through ONNX, so the page stays usable. It is slower than a cloud service, and on a laptop long files can take longer than the recording itself.
From words to cues
Timed words are grouped into cues: a new cue starts after a pause of 0.8 seconds, after a sentence ends, at 42 characters or at six seconds, whichever comes first. Each cue can be reworded and retimed in the list; use the number to jump the player to it.
new cue when gap > 0.8 s, or length > 42 characters, or duration > 6 s, or the word ends a sentence
The model download
The first run downloads the model from the Hugging Face hub: about 80 MB for Base and about 250 MB for Small. The browser keeps it in its cache, so later runs start straight away. Only the model files travel over the network; your audio never does.
