Turn-taking

What it is

Turn-taking is the umbrella behaviour built from several mechanisms: voice activity detection, endpointing, end-of-turn detection, barge-in and latency management. Humans do this without thinking, using pitch, pace, grammar and a lifetime of practice, and gaps between speakers in ordinary conversation are typically a fraction of a second. A voice agent has to approximate the same thing from a compressed telephone signal.

Good turn-taking is invisible. Bad turn-taking is the first thing a caller notices: being cut off, being talked over, long silences, or an agent that answers a question the caller had already moved past.

Why it matters when buying

Turn-taking is the property that separates voice from chat and separates production-grade agents from demos. It cannot be fixed by a better prompt. It depends on the architecture (whether the components exchange timing information or run as a strict pipeline), on tuning for the telephony path you use, and on how the agent behaves during slow tool calls. Vendors that demo on a headset in a quiet room, with the demo operator's own steady speech, are not demonstrating turn-taking.

What to ask

Run the demo from a mobile phone, from a room with some background noise, with a caller who has not seen the script. Interrupt, pause mid-sentence, give one-word answers, and read a number in groups. Score each behaviour pass or fail. Ask the vendor how each mechanism is implemented and which are tunable per deployment. The demo guide has the full trap list.