For a voice platform · conversation · 2025
What happens when you interrupt an AI mid-sentence — and can it interrupt back?
Turn-taking as design material — overlap isn’t an error, it’s a state with rules.
Today most voice agents take strict turns — interrupt one and it either talks over you or loses the thread.
Why now real-time voice agents are arriving fast, and clumsy turn-taking is what still breaks the illusion.
Use cases Phone support, receptionists, in-car assistants, voice ordering.
Where it went · 2026 I built this in late 2025 — and the frontier has since landed right here. In July 2026 OpenAI shipped GPT-Live, a full-duplex voice model that listens and speaks at once and, when a turn needs deeper reasoning, hands off to GPT-5.5 running in parallel so the agent keeps talking instead of going quiet. The two things this prototype was poking at — interrupting mid-sentence, and not freezing while it thinks — are now in production. What I’m watching next is turn-taking etiquette: when should an agent yield, hold, or talk back?
Interactive demo ↗The question
Real talk overlaps. People cut in, murmur “mm-hm”, trail off. Most voice agents treat all of that as a mistake to suppress.
The approach
Make turn-taking visible. Every interruption, back-channel and pause becomes a named state on a strip you can watch — and afterwards, a two-track trace you can read.
The feature
Interrupt by button, by keyboard, or with your own voice level. The agent yields, holds or resumes by rules you can switch on and off.
Function
The turn-taking states
idle, speaking, listening-while-speaking, yielded, resumed, holding — the conversation as a machine you can see.
Function
The two-track trace
Afterwards the whole exchange reads as two tracks — where speech overlapped, and which rule resolved it.