Efficient Causal End-of-Turn Prediction for Real-Time Spoken Dialogue Systems

A voice agent must decide, many times per second, whether the person it is talking to has finished their turn: answering early cuts the speaker off, and waiting makes the conversation feel slow. The strongest learned detectors run on servers, so most deployed agents still rely on silence timers. We show that this trade-off can be solved on the device. turn-1-mini is a strictly causal, audio-only end-of-turn model with 5.3M parameters (5.4 MB in 8-bit form) that runs in 1.3 ms per 160 ms of audio, under 1% of a single laptop CPU core, and needs no transcript. On Sesame's TurnBench official test set it ranked first of 23 entries on end-of-turn detection (recall 0.975 at a 0.075 false-positive rate, 595 ms median latency) and second of 20 on interruptions. On LiveKit's eot-bench (English) it has the lowest false cut-off rate (22.6% at 300 ms) and the lowest latency at a 5% false cut-off rate (849 ms) of every downloadable model, while being about 530 times smaller than the server-class model it was distilled from. Against voice-activity rules it commits interruption decisions 572 ms earlier at higher recall. Weights, code and ONNX streaming graphs are released under Apache-2.0.