← Back to glossary
+Suggest a term
Concept·AI Models & Capabilities·Added today

Full-duplex voice

Also known as: full-duplex AI voice, duplex voice model, simultaneous listen-speak

A voice AI architecture where the model listens and speaks at the same time, rather than waiting for the user to finish before responding. This enables natural interruptions, overlapping speech, and conversation flows that feel more like a phone call than a voice assistant.

Most voice AI systems before 2026 were turn-based: the model detects silence, processes the audio, generates a response, plays it back, then waits again. This is fine for simple commands but breaks down in real conversation, where people interrupt, overlap, trail off, and recover mid-sentence. Full-duplex voice architecture changes the fundamental loop: the model processes incoming audio continuously, even while it is speaking, and decides many times per second whether to keep talking, pause, acknowledge, or hand off to a background model for a harder task.

The distinction matters for builders because it changes what voice agents can realistically do. Turn-based systems work for structured dialogs, like booking a table or checking an account balance. Full-duplex systems make it practical to build agents that can handle messier, open-ended conversations: sales calls, support escalations, real-time coaching, and ambient assistants that stay present without being intrusive.

GPT-Live, OpenAI's full-duplex voice model family launched in July 2026, is the highest-profile consumer example of this architecture to date. The pattern is distinct enough from prior voice AI that 'full-duplex voice' is entering builder vocabulary as its own design consideration, separate from the underlying models that implement it.

This definition is AI-generated and refreshed weekly. It may contain inaccuracies. Use your own judgment, especially for production decisions.
Related terms
GPT-LiveText-to-speechMultimodal modelVoice cloningAlways-on agent