concept

Duplex conversation

Duplex conversation is the realtime-voice property where both parties (user + AI) can speak + listen simultaneously — supports natural interruption, backchanneling ('uh-huh'), and overlapping speech. The bar for voice agents to feel human in 2026.

Half-duplex voice (one party talks at a time, walkie-talkie style) feels robotic. Duplex flips this: continuous bidirectional audio, both parties can speak / interrupt / backchannel naturally. Engineering: audio streaming both ways simultaneously (WebRTC handles this natively), [[barge-in]] detection (stop AI playback when user starts speaking within ~100ms), backchanneling ('mm-hmm', 'right') without breaking the user's turn, turn-taking models that predict who should speak next. Realtime APIs (OpenAI, Gemini, ElevenLabs Conversational) ship duplex out of the box. The difference between voice that feels natural and voice that feels like calling a customer service IVR.

When to use duplex conversation

Common mistakes

FAQ

What is duplex conversation?

Duplex conversation is the realtime-voice property where both parties (user + AI) can speak + listen simultaneously — supports natural interruption, backchanneling ('uh-huh'), and overlapping speech. The bar for voice agents to feel human in 2026.

When should I use duplex conversation?

Voice agents in customer-facing apps. Phone agents, voice assistants.

What are the most common mistakes with duplex conversation?

Building half-duplex voice apps in 2026 — feels dated immediately. Disabling barge-in for 'clean audio' — kills the natural feel.

Last updated: 2026-06-01. Raw markdown: https://promtable.com/glossary/duplex-conversation.md.