Semantic Turn Detection
for Voice AI Systems

This study investigates how voice AI agents decide when you've finished speaking. You'll have a short conversation with an AI across three different conditions - each using a different strategy for detecting the end of your turn.

Select a condition to begin
A
VAD Baseline

The AI responds after a fixed period of silence. No language understanding is used; the system simply waits for the mic to go quiet.

Fixed 1.0s threshold Silence-only
B
VAD + Syntactic Scorer

Silence detection is combined with a lightweight language model that scores whether your utterance forms a complete thought before firing.

EOS scorer Completeness gate
C
Full Ensemble

Combines intent classification, a dynamic silence threshold, an EOS score, and an LLM turn-completion scorer into a weighted ensemble decision.

Intent-aware Dynamic threshold Ensemble
Survey Complete feedback form

Semantic Turn Detection

Deliberative Voice AI Study

A
VAD Baseline
Responds after a fixed 1.0s silence window. No language understanding - silence alone determines the turn boundary.
Fixed 1.0s threshold Silence-only
The AI will ask you
I'm here to help you think through a decision. Tell me about something you've been going back and forth on - walk me through both sides as you see them.
idle
Click the mic to begin
Rolling transcript (u_t)
Ensemble scoring
EOS
LLM scorer
Combined