Request a new model

Tell us which model you want measured. We review every request.

Turn-taking · end-of-turn detection

01
Smart Turn v3.2
Pipecat · Whisper-tiny
audio · prosody
94.0%
89.0%
1.0%One false cutoff in 100 incomplete clips.
32msRuns in parallel with STT — no transcript wait.49 p90
02
LiveKit Intl
turn-detector v0.4.1
text · transcript
87.0%
86.0%
12.0%
29msWaits for the STT transcript (~0–125ms) before running.57 p90
03
Turnsense
SmolLM2-135M
text · transcript
86.0%
88.0%
16.0%
127msPads every input to 256 tokens regardless of length — slowest as-shipped.193 p90
04
LiveKit EN
turn-detector v1.2.2
text · transcript
82.0%
87.0%
23.0%
5ms24 p90
05
VAD + silence timer
baseline
audio · energy
46.9%
100%
100%A silence timer ends every turn — the floor semantic detection has to beat.
~1ms
ModelAccuracy syntheticFalse-cutoff synthetic
Turnsensetext98.8%0.0%
LiveKit Intltext90.6%6.2%
LiveKit ENtext79.2%26.2%
Smart Turnaudio62.4%57.7%
Acoustic cueComplete ENDIncomplete WAITWhat it means
Final F0 slope-119.7 Hz/s+42.9 Hz/scomplete falls, incomplete rises/holds
Final energy slope-14.9 dB/s-2.5 dB/scomplete tapers off

FAQ

What is end-of-turn detection in a voice agent?
It is the model's decision that the caller has finished speaking rather than paused mid-thought. On 200 real human clips the best detector, Smart Turn v3.2, got 94.0% of END-vs-WAIT decisions right; a plain VAD silence timer scored 46.9%.
Which turn detection model is the most accurate?
Smart Turn v3.2 measured 94.0% accuracy with a 1.0% false-cutoff rate. The best text-based detector, LiveKit Intl turn-detector v0.4.1, measured 87.0%.
What is a false cutoff in turn detection?
An incomplete turn wrongly ended: the agent talks over the caller mid-thought. Among semantic detectors, measured rates span 1.0% (Smart Turn v3.2) to 23.0% (LiveKit EN turn-detector v1.2.2).
Is VAD good enough for turn detection?
No. A VAD-plus-silence-timer baseline ends every turn at the timer, so it measured 100% false-cutoff and 46.9% accuracy, the floor semantic detection has to beat.
Should I use audio-based or text-based turn detection?
On real human audio the audio-prosody model led: Smart Turn v3.2 at 94.0% against 87.0% for the best text detector, LiveKit Intl. On synthetic TTS speech the ranking flips: Smart Turn drops to 62.4% while text-based Turnsense scores 98.8%.
How much latency does turn detection add to a call?
Warm-CPU inference measured 5ms (LiveKit EN turn-detector v1.2.2) to 127ms (Turnsense). Audio models run in parallel with STT; text models additionally wait on the transcript.