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