OpenAI rebuilt ChatGPT Voice to listen and speak at the same time

OpenAI rebuilt the voice mode behind ChatGPT so the model can listen and speak at the same time. The new system, called GPT-Live, is the company's third-generation voice architecture. The post positions it as the foundation for a newly launched ability in the ChatGPT desktop app to control your computer and coordinate agents.

The change, laid out in an engineering post by OpenAI's Justin Uberti and Zahan Malkani, is small to describe but large in practice. Earlier ChatGPT Voice worked turn by turn: a small gating model listened to the user's audio, decided when the turn was over, and only then handed the audio off to the larger model for a reply. The reply came back as a separate audio blob. That pipeline left two familiar failure modes: the gate sometimes cut the user off mid-sentence, other times it waited too long and the assistant felt sluggish. The tradeoff was structural, not a tuning problem.

GPT-Live collapses that two-step pipeline. The voice model is full-duplex, meaning it streams audio in both directions at once, and the gating model is no longer sitting in the audio path deciding when the AI can speak. The cut-off-vs-sluggish tradeoff is, in OpenAI's framing, gone with it.

For harder questions, GPT-Live does not have to do all of its own thinking in real time. It delegates to a frontier model on a separate asynchronous path. The post names GPT-5.5 as one such model, and the design is meant to keep that delegation from interrupting the conversational flow. The clean line between the voice path and the model doing the deeper work is the second architectural choice the post leans on.

The post also sketches a four-layer stack the new system optimizes across: stateful inference (the model keeps memory of the conversation across turns); dynamic context management (deciding what context to keep loaded); asynchronous delegation (offloading reasoning to a heavier model when needed); and protocol-level optimization (how audio, text, and tool calls move between the model and the app). The whole system was reworked over six months, and the post treats the change as much about plumbing as about the model.

For a reader outside the voice-AI beat, the portable mental model is simple. Older voice assistants are turn-based: speak, wait, get a reply. Full-duplex means the assistant can hear you while it is still answering, and can interject, hesitate, or course-correct in a way turn-based stacks cannot. That difference is what the industry has been chasing for years, but it has a known failure mode: the model talking over you, or hallucinating mid-sentence because it is no longer gated by a turn detector.

That second half of the tradeoff is the part OpenAI's post does not address. The post frames the removal of the turn detector as pure progress on naturalness. It does not show latency numbers, error rates, or independent test results. The harder question is how the new system handles a user who starts talking while the assistant is mid-answer, and how it compares in latency and naturalness to OpenAI's prior ChatGPT Voice or competitors like Hume, ElevenLabs, and Sesame. Those are open questions for a reporter chasing the story.

The architectural claim is concrete and on the record. The user-experience claim is the company's, and not yet independently verified.

The ChatGPT desktop app is the first place the new system is being used to coordinate agents and control a computer, so any failure of full-duplex voice in that surface (the assistant talking over a user mid-command, for example) will be visible in normal use, not just in a benchmark. The post promises to cover each of the four layers in follow-up engineering pieces; those are where the latency, error, and comparison data will show up first.