Turn detection · Open weights · Apache-2.0

turn-1 and turn-1-mini.

Know when the speaker has finished, so your agent neither cuts people off nor leaves them waiting.

By Kollaikal Rupesh · research@p99lab.com

Models

Two models

Small · streamingReleased turn-1-mini Our own turn-detection model: our architecture, our weights and our training, with no third-party weights inside. Ranked first on end-of-turn on Sesame's TurnBench. Parameters5.3M Size21 MB · 5.4 MB with 8-bit weights ONNX build7 MB, runs without PyTorch Speed1.3 ms per 160 ms step, one CPU core (Apple M4 Pro) huggingface.co/p99lab/turn-1-mini

Larger · serverReleased turn-1 Our larger end-of-turn model for servers. Second on the public eot-bench English benchmark (point estimates). Runs onServers Benchmarkeot-bench English, 2nd LicenceApache-2.0 Plug-insPipecat · LiveKit huggingface.co/p99lab/turn-1

Results · Sesame TurnBench

TurnBench

turn-1-mini is ranked first on end-of-turn and second on interruption on Sesame's TurnBench (held-out test set). The only open-weights model in the top five on either task. Leaderboard page

1stof 23End-of-turn

2ndof 20Interruption

0.975End-of-turn recall

595msMedian end-of-turn latency

Task	Rank	Recall	False-positive rate	Latency (median)	Latency p10	Latency p90
End-of-turn	1st of 23	0.975	0.075	595 ms	156 ms	1,243 ms
Interruption	2nd of 20	0.973	0.048	551 ms	295 ms	1,576 ms

Sesame TurnBench · turn-1-mini · held-out test set. 116 conversations, scored by Sesame with their scorer. The leaderboard ranks by recall. Recall: share of true events detected. False-positive rate: lower is better. Latency: delay after the event.

Results · LiveKit eot-bench · English

eot-bench

turn-1 is second on the public eot-bench English benchmark, behind LiveKit Turn Detector v1 and ahead of Deepgram Flux and AssemblyAI on all four measures (point estimates), and needing about half the wait of OpenAI GPT Realtime 2 on the two latency measures (point estimates). turn-1-mini is ahead of every downloadable model on the same benchmark, including Smart Turn v3.2 and LiveKit's v1-mini (point estimates). Never trained, selected or thresholded on the benchmark by us. Every scored run is logged.

Loading results…

LiveKit eot-bench · English. Selected rows. Lower is better in every column. False cut-offs: how often the model ends a turn the speaker had not finished, at a fixed wait. Latency: how long the model waits before ending a turn, at a fixed cut-off rate. Blank cells: no score on the board for that measure. Our models are highlighted.

How they are built

turn-1-mini is entirely our own: a 5.3M-parameter streaming model that we designed and trained. It contains no third-party weights.

turn-1 is a turn-detection model by p99lab, built on open speech encoders. The turn-detection model, its data pipeline and its training are ours.

Licence

turn-1 and turn-1-mini are released under Apache-2.0. The speech encoders inside turn-1 keep their own licence.

TurnBench write-up: github.com/P99Lab/turn-1-mini