pardon / audio labON YOUR DEVICE

REALTIME TURN DETECTION

A pause—or the
end of a thought?

Speak naturally. Watch the model listen for a finished turn.
Your microphone audio stays in this browser.

Start listening

Load the model once, then try your microphone or the spoken example.

One-speaker mode listens to the first channel and supplies silence for the other. No retraining or speaker separation.

Speaker 1

Ready
—turn-complete score
Speech activity —

Keeping up

—processing lag

— per decision · 0.00 s heard

Decisions every 160 ms. If lag grows, close other GPU-heavy tabs.

The last 12 seconds

● Turn score― Speech activity
Dashed line: fixed turn threshold (82.1%). A turn fires only after observed speech, while that speaker is quiet.

Detected turns

Finished turns will appear here.

    About this experiment & device check

    FP32 streaming Parakeet encoder + Pardon turn head + Silero speech activity. Encoder inference uses WebGPU with GPU-resident recurrent caches; small head/VAD operations run locally on WebAssembly. Some encoder operations may also use CPU fallback.

    One-speaker mode is a silent-companion baseline. Its accuracy has not been established on single-speaker tasks. Two-speaker mode requires one speaker on each stereo channel; it does not separate mixed voices. Browser resampling and microphone processing can differ from the research pipeline.

    Research preview: numerical conversion checks are not a TurnBench score. No transcript, server inference, microphone upload, or benchmark submission. Model files are public and cached by your browser when storage is available.

    Not run on this device.

    Base model & license · Silero VAD

    Attribution notice · Model license