Turn-taking playground

A streaming turn-taking model running entirely in your browser (WebGPU, or CPU as a fallback). Feed it two speakers, from your microphone or audio files, and watch it decide, every 80 ms, when a speaker has finished their turn and when a speaker is taking the floor. Nothing is uploaded: audio stays on your device.

Backend

The model is about 225 MB and is cached after the first load. Start also loads it.

Sources

One source per speaker One stereo file (left = A, right = B)

Speaker A

Speaker B

Background noise and options

Background audio mixed into speaker A (TV, music, other talkers; loops) Background level Play files through the speakers (the microphone is never played back) Browser noise suppression and auto gain on the microphone

A end of turntakes the floor

B end of turntakes the floor

speaking input level end-of-turn score take-the-floor score dashed: thresholds; markers: committed events

End-of-turn threshold 0.64 Take-the-floor threshold 0.18

Backend: not loaded
Time per call: –
Real-time factor: –
Latency in pipeline: –
Stream time: –

About

The model is a causal FastConformer encoder (NVIDIA stt_en_fastconformer_hybrid_large_streaming_multi, CC-BY-4.0) feeding a small cross-speaker turn head trained on two-channel conversations, with background-speech augmentation. It sees 80 ms frames with no lookahead. The default thresholds are the operating points on the TurnBench dev set (end of turn: recall 0.93 at 10% false fires; take the floor: recall 0.98 at 9%). Events are committed on a rising edge above the threshold with a 0.5 s refractory, as TurnBench scores them.

Two scores per speaker: end of turn rises when that speaker has stopped and is handing over the floor (not a mid-sentence pause); take the floor rises when that speaker starts a real turn, not a backchannel like "mm-hm". The model is English-only and a research preview, for non-commercial use.