Full-duplex voice AI is a voice assistant that can listen and speak at the same time, the way two people do on a phone call, instead of taking strict turns. That single change is why newer voice assistants feel like a conversation rather than a series of walkie-talkie exchanges: you can interrupt mid-sentence, correct yourself, or jump in with a quick “wait, actually—” and the assistant reacts immediately instead of finishing its rehearsed answer first.
What “duplex” means
The term comes from telecommunications. In a duplex system, two parties can send and receive at the same time (full-duplex, like a phone call) or only one direction at a time (half-duplex, like a walkie-talkie, where you have to say “over” before the other side can talk). Early AI voice assistants — including the first versions of ChatGPT’s spoken mode — were effectively half-duplex: you spoke, the app waited for you to stop, then it thought and replied. Full-duplex voice AI moves the underlying model to the phone-call model instead.
From a three-step pipeline to one continuous stream
Most early voice assistants were built from three separate stages chained together: speech recognition turns your audio into text, a language model reads that text and drafts a reply, and a text-to-speech engine reads the reply out loud. Each stage has to finish before the next one starts, which adds up to a noticeable pause, and none of the stages know what’s happening in the other two — so the system can’t react while you’re still talking.
Full-duplex systems replace that relay race with a single model that processes audio in and produces audio out continuously, over an open connection, rather than in one-shot exchanges. Because the model is listening the whole time, it can detect when you start talking over it — commonly called “barge-in” — and stop or adjust immediately instead of finishing a sentence you’ve already cut off. OpenAI’s GPT-Live family, released for ChatGPT’s voice mode, works this way and hands off harder reasoning or search tasks to its main language model so the voice layer itself can stay fast and responsive. OpenAI’s developer-facing Realtime API, now running on the gpt-realtime-2.1 model, gives the same audio-in/audio-out streaming to outside apps over a persistent connection. Other labs are building similar systems — NVIDIA’s Nemotron VoiceChat, for instance, is a full-duplex speech-to-speech model aimed at the same kind of real-time conversational use.
Why it matters
The practical effect is that voice stops feeling like a form you fill out one field at a time and starts feeling like talking to a person. That matters for anything where natural back-and-forth is the point: customer support calls, language practice, brainstorming out loud, or hands-free use while driving or cooking, where waiting out a long, uninterruptible answer is the main reason people give up on voice assistants. It also matters for accessibility — people who find typing slow or difficult get a much closer approximation of a real conversation. The tradeoff is that a system this good at mimicking natural conversational rhythm also raises the bar for voice-based scams and impersonation over the phone, which is part of why deepfake detection has become its own field.
How to try it
In the ChatGPT app, the newer “Live” voice option listens and speaks at once by default; users who prefer the older back-and-forth style can switch to “Separate mode” in the app’s voice settings (see OpenAI’s ChatGPT Voice guide). Developers building their own voice apps can connect directly to OpenAI’s Realtime API, which streams audio over a WebSocket, WebRTC, or SIP connection; usage is billed per audio token rather than a flat fee, and current rates are listed on OpenAI’s API pricing page.
In the news
OpenAI launched GPT-Live as a full-duplex upgrade to ChatGPT’s voice mode, and followed it with GPT-Realtime-2.1, a developer-facing model built for the same real-time, interruptible audio streaming with lower latency.
FAQ
Is this the same as Siri or Alexa’s voice mode? Not usually. Most older voice assistants still process a command, respond, then wait for the next one — closer to half-duplex. Full-duplex systems are a newer generation built to be interrupted and to react while you’re still speaking.
Does it understand tone of voice, not just words? Because the model processes raw audio instead of a text transcript, it can pick up on some cues a transcript would lose, like pacing or emphasis. It’s still not full emotional understanding, but it’s closer than a text-only pipeline.
Do I need to pay for it? ChatGPT’s voice mode, including Live, is available to logged-in users on the free tier as well as paid plans, with usage limits varying by plan. Building your own app on the Realtime API is billed separately, per audio token.
Sources: OpenAI’s Realtime API documentation and ChatGPT Voice help article; Wikipedia’s entries on duplex communication and speech recognition.