voxturn

Open turn-taking model for voice agents

Knows when you're
done talking.

voxturn listens to the last 8 seconds of a conversation and answers the questions a voice agent actually has — finished? about to go on? just a cough? — in a single forward pass.

▶ Hear the difference Get the model
11.5%false cut-offs at a 300 ms budget
English eot-bench
19.9%14-language mean
best among open models
8.3 msper decision
on one GPU
2.7×fewer turns cut off
than Pipecat's default

Live replay

Same framework. Same VAD. Different turn model.

Real user turns from eot-bench, streamed through Pipecat's Silero VAD — once with Pipecat's default Smart Turn v3, once with voxturn. Press play and watch who jumps in too early.

Agent
User
voxturn P(done) Smart Turn v3 P(done) mid-turn pause user finished
0.0 s
Pipecat + Smart Turn v3 default
listening…
Pipecat + voxturn
listening…

The same forward pass also answers (voxturn, at the playhead)

How it works

A speech LLM that doesn't talk — it decides.

01

Questions, not labels

Every decision is a multiple-choice question for Qwen3-ASR-0.6B. No text is generated: a tiny head scores the hidden state of each option, and a softmax gives calibrated probabilities.

eotresumestate4intentinvalid
02

One pass, many answers

Questions sit after a shared audio prefix, each in its own block of the attention mask — so one pass equals asking them one by one. Four questions: 9.0 ms instead of 33.2 ms.

03

Learned from real pauses

Real phone calls and face-to-face talk, with the decision time drawn anywhere inside each pause. voxturn grows more certain as a silence goes on — so we keep asking every 250 ms and reply the moment it's sure.

eot-bench

Commercial-grade in English. Best open model in 14 languages.

False cut-offs (the agent replies while the user is only pausing) at each mean-latency budget. Lower is better. Hover the chart.

English: false cut-offs vs latency budget

Every language, 300 ms budget (open models)

Pipecat45.3%16.7%turns cut off · Smart Turn v3 vs voxturn
LiveKit18.0%9.3%turns cut off · local v1-mini vs voxturn

Drop it in

Three lines to stop interrupting people.

from huggingface_hub import snapshot_download
import sys; sys.path.insert(0, snapshot_download("AustinXiao/voxturn-v1"))
from voxturn import VoxTurn

model = VoxTurn.from_pretrained("AustinXiao/voxturn-v1")
model.predict(audio, sample_rate=16000, questions=["eot", "resume", "invalid"])
# {'eot': {'yes': 0.93, 'no': 0.07}, 'resume': {...}, 'invalid': {...}}
from pipecat_voxturn import VoxTurnAnalyzer
from pipecat.turns.user_stop import TurnAnalyzerUserTurnStopStrategy
from pipecat.turns.user_turn_strategies import UserTurnStrategies

strategies = UserTurnStrategies(stop=[TurnAnalyzerUserTurnStopStrategy(turn_analyzer=VoxTurnAnalyzer())])
from livekit_voxturn import VoxTurnDetector
from livekit.agents import AgentSession
from livekit.plugins import silero

session = AgentSession(vad=silero.VAD.load(min_silence_duration=0.3),
                       turn_handling={"turn_detection": VoxTurnDetector()})