AustinXiao/voxturn
Audio Classification • Updated • 464 • 8
VoxTurn
Open turn-taking models for voice agents: end-of-turn, barge-in and backchannel detection. voxturn (multilingual) + ShadowMe (speaker-aware).
Note voxturn v1: 0.8B multilingual turn model (Qwen3-ASR-0.6B). One pass answers end-of-turn, resume timing, backchannel vs. interruption, and speech vs. noise. eot-bench EN 11.5% false cut-offs @300 ms; 19.9% mean over 14 languages; ~8 ms per decision. Pipecat & LiveKit plugins.
AustinXiao/voxturn-personal-turn-detector
Audio Classification • Updated • 111Note ShadowMe (preview): speaker-aware sibling (2.1B, Qwen3-ASR-1.7B + ReDimNet-B3). Enrolls ~10 s of the user's voice and decides end-of-turn and barge-in for that user only, ignoring bystanders. eot-bench EN 10.4% @300 ms. English, Pipecat.
voxturn
🎙Open turn detector - has the user finished talking?
Note Interactive demo + hosted API (ZeroGPU): upload or record a turn and see P(end of turn) every 100 ms; compare ShadowMe with voxturn v1. Callable via gradio_client, no GPU needed.