Barge-in and end-of-turn detection: what makes a voice agent feel human
The two turn-taking skills that separate natural voice agents from talking menus: yielding instantly when interrupted, and knowing when the caller is done.
16 June 2026 · 2 min read · by the Dayl team
Key takeaways
- Barge-in, stopping instantly when the caller speaks, is the single strongest signal that an agent is listening rather than playing audio.
- End-of-turn detection by silence alone forces a bad trade: interrupt the caller's pauses, or add dead air to every turn.
- Semantic end-of-turn judges whether the utterance is linguistically complete, cutting response delay without cutting callers off.
- Noise and echo are turn-taking saboteurs: false barge-ins from background sound make an agent feel broken.
Interruption is a feature, not an error
Real callers interrupt constantly, to correct, to redirect, to answer early. A natural agent stops speaking within a fraction of a second of the caller starting, discards its now-irrelevant plan, and listens. An agent that keeps talking over the caller reveals itself as a recording; nothing else it does well will recover that impression.
Clean barge-in requires echo discipline: the system must distinguish the caller's voice from its own synthesized speech leaking back down the line, or it will interrupt itself. This is why echo cancellation and voice isolation sit upstream of turn-taking logic.
The silence-threshold trap
Classic end-of-turn detection waits for N milliseconds of silence. Set N short and the agent barges into the caller's mid-sentence pauses, devastating in Arabic, where a caller reciting an address naturally pauses between segments. Set N long and every turn inherits that delay as dead air. There is no good value of N; the mechanism itself is wrong.
Semantic end-of-turn replaces the timer with a judgment: given the words so far, is this utterance complete? 'I want to change my delivery to' is clearly unfinished regardless of pause length; 'yes, tomorrow afternoon works' is clearly done. Models making this call react in a few hundred milliseconds on complete turns while patiently waiting through mid-thought pauses.
Turn-taking under real-world audio
Every turn-taking mechanism degrades with dirty audio: a door slam reads as speech onset, background chatter as the caller resuming. Production systems therefore treat turn-taking as a pipeline property, cleaned audio in, semantic completion judgment, echo-safe interruption, and instrument every false barge-in and delayed response as a metric, because turn-taking quality is measurable, not anecdotal.
Frequently asked questions
Within a few hundred milliseconds of genuine caller speech. Slower feels like talking to a broadcast; the caller repeats themselves and frustration compounds.
Sources & further reading
Go deeper