Full-Duplex Voice

What It Is

Full-duplex voice is the capability that lets an AI assistant hear you and talk to you in the same moment, the way two people actually converse. Most voice assistants you have used are half-duplex: they wait for you to finish, process what you said, then reply, and while they are speaking they are effectively deaf. Full-duplex removes that turn-taking wall. The system keeps its microphone open while its own voice is playing, so you can cut in, correct it halfway through a wrong answer, or say “no, the other one” without waiting for it to stop.

The term moved into everyday reach this week when OpenAI introduced GPT-Live, a mode inside the new ChatGPT desktop app that lets the model listen and speak simultaneously for a more natural back-and-forth. The difference sounds small on paper and feels large in use. A half-duplex assistant makes you manage the conversation around its rhythm. A full-duplex one lets you keep your own.

How It Actually Works

The hard part is not the speaking, it is doing both at once without the system talking over you or mishearing its own voice as your input. Three things have to run continuously and in parallel. A streaming speech recognizer transcribes your words as they arrive, not after you pause. A turn-taking model constantly judges whether you are interrupting, agreeing, or just saying “mhm,” and decides in real time whether to keep talking or yield. And echo cancellation strips the assistant’s own audio out of the incoming microphone feed so it does not react to itself.

All of that has to happen fast enough that the reply feels immediate, which puts heavy weight on low-latency inference. A model that reasons brilliantly but answers a second and a half late breaks the illusion of conversation entirely. This is why full-duplex is as much an engineering achievement in serving speed as it is a language achievement. The intelligence was mostly there already in the underlying large language model; the new work is delivering it inside the tight timing budget human speech expects.

Why It Matters Right Now

Voice is the interface that finally lets an assistant ride along while your hands and eyes are busy, in the car, in the kitchen, on a walk, on a factory floor. Half-duplex voice never quite worked for that because the stilted turn-taking made it faster to just pull out your phone. Full-duplex closes enough of that gap that voice becomes a serious primary interface rather than a novelty, which is why every major lab is now racing to ship it rather than treating it as a checkbox feature.

It also changes what a voice assistant can be paired with. Once the model can hold a fluid spoken exchange, it becomes far more natural to let it drive actions while you talk, narrating what it is doing and taking correction on the fly. That is where voice starts to overlap with computer use and with the broader idea of an AI agent: the conversation is no longer just questions and answers, it is a running dialogue with something that can also act.

How TWO Uses It

TWO treats full-duplex voice as a genuine interface upgrade and a discipline trap at the same time. The upgrade is real: for the right task, talking to a capable model that can be interrupted is meaningfully faster than typing. The trap is that a smooth, human-sounding, always-listening assistant lowers your guard precisely when it should stay up. A microphone that never fully closes is a microphone that never fully closes, and the friendlier the voice, the easier it is to forget that.

Scott’s Take: The more natural the machine sounds, the more deliberately you have to decide what you actually want it to hear.

The concrete operator decision is where you let it run open. Turn full-duplex voice on for genuinely hands-busy, low-stakes work, drafting out loud, thinking through a plan, walking through a checklist. Keep it off, or keep it half-duplex and push-to-talk, for anything involving sensitive material or private conversation happening in the same room. The skill is not adopting or refusing the feature. It is knowing, per task, whether the convenience of an open mic is worth the reach of one.

What to Watch Next

The signal to watch is not smoother speech, it is where the listening actually happens. The moment worth noticing is when full-duplex voice runs entirely on your own device, so the open microphone feeds a local model rather than a stream to a vendor’s servers. On-device full-duplex would resolve most of the privacy tension by keeping the always-listening part in your own hands. Until then, treat every natural-sounding, ever-attentive assistant as exactly what it is: a real convenience with a standing invitation to overhear.