On-device turn detection
Done speaking, or just pausing? Runs entirely in your browser.
Model says
Pick an example, upload a file, or record
No audio yet
P(complete)
0.000
Load some audio to get a reading.
The model hears only the audio. It judges whether your speech sounds finished, not whether you stopped recording. Silence after you stop is not evidence that you were done. It gets this wrong sometimes, and it leans toward waiting, so a finished sentence can still read as "Still speaking".
Labelled examples
The pill is the true label. Run one and the model verdict appears beneath it.
The model reads the last 8 s. Shorter clips are padded with silence on the left.
P(complete) recomputed on growing prefixes of the clip, 0.24 s apart, with the selected model.