October 20, 2026 · 8 min read
What Is Full-Duplex Voice AI? Why AI Calls Finally Feel Like Real Conversation
Full-duplex voice AI listens and talks at once, so you can interrupt it like a person. How it works, GPT-Live-1.1, and what it means for AI calls.
Think about the last good phone call you had with someone you love. You laughed over each other. They said "mm-hm" while you told the story. You cut in with "wait, what?" and they stopped mid-sentence to explain.
Until recently, talking to an AI felt nothing like that. You spoke, waited, and then it spoke, like passing a walkie-talkie back and forth. That is changing fast, and the technical name for the change is full-duplex voice AI.
On October 15, 2026, OpenAI released GPT-Live-1.1, the latest version of its full-duplex voice model for ChatGPT. It is a good moment to explain what full-duplex actually means, why it makes AI calls feel so much more human, what the research says about the timing of real conversation, and what to look for in an AI companion that calls you.
The short answer: Full-duplex voice AI can listen and speak at the same time, like a phone call, instead of waiting for you to finish a turn, like a walkie-talkie. That lets it handle interruptions, pauses and "mm-hm" moments naturally. OpenAI's GPT-Live-1.1, released October 15, 2026, is one example. It makes AI voice chat feel more like talking to a person, but it does not make the AI a person, and its own safety report shows some new weak spots.
Half-duplex vs full-duplex, in plain English
The terms come from telecoms:
| Half-duplex | Full-duplex | |
|---|---|---|
| Everyday example | Walkie-talkie | Phone call |
| Who can talk | One side at a time | Both sides at once |
| Interrupting | Usually breaks the flow or is ignored | Handled like in real life |
| Short "mm-hm" replies | Often treated as a new turn | Can be heard as listening cues |
| Feel | Polite but stiff | Lively and natural |
Most older voice assistants were effectively half-duplex. They used a pipeline: turn your speech into text, send the text to a language model, then turn the reply back into speech. Each step waits for the one before it, and the system has to guess when you have finished talking.
Why the old way felt robotic
Researchers at the French lab Kyutai described the problem clearly in their arXiv paper (a preprint) on Moshi, an open full-duplex speech model. Pipeline systems, they wrote, add "a latency of several seconds", lose emotion and other non-written cues between steps, and rely on rigid turns.
Real conversation is not like that. The Moshi team notes that overlapping speech and interruptions make up a meaningful share of spoken time in natural dialogue, and that people constantly use small sounds like "okay" and "I see" to show they are listening. A system that only waits for silence will either cut you off or keep you hanging.
Moshi itself reported a theoretical latency of 160 milliseconds and about 200 milliseconds in practice. That matters because humans are fast too. A study of conversations in ten languages published in PNAS by Tanya Stivers and colleagues found that answers to yes-or-no questions most often arrive within 200 milliseconds of the question ending. We dig into why that timing matters on calls in how AI video calls work.
What full-duplex models do differently
A full-duplex model keeps listening while it talks. In OpenAI's words from the GPT-Live-1.1 system card, these models "can listen and respond continuously instead of waiting for a clearly defined turn to end", which means "they can follow pauses, interruptions, and changes in pace, and decide in the moment whether to respond or keep listening".
In practice, that unlocks things like:
- Barge-in: you can interrupt and it stops or changes course.
- Thinking pauses: you can say "hmm, let me think" without it jumping in.
- Backchannels: it can tell the difference between "mm-hm" (keep going) and a real question.
- Pace matching: it can follow you when you speed up with excitement or slow down when you are tired.
GPT-Live-1.1: what OpenAI released
According to the system card dated October 15, 2026:
- Two versions. GPT-Live-1.1 is the default voice model for ChatGPT Plus and higher plans. GPT-Live-1.1 Quip is the default for other users.
- A two-model system. OpenAI says GPT-Live "is actually a two-model system", with a voice model working alongside a "thinking" model.
- Safety monitors. OpenAI says automated monitors "may interrupt and end the call" when potentially harmful conversations are detected.
The ChatGPT Voice help page adds the practical details:
| Plan | Live voice in chat (rolling 24 hours) |
|---|---|
| Free | Limited access to GPT-Live-1.1 Quip |
| Go | 3 hours with Quip |
| Plus | 3 hours with GPT-Live-1.1 |
| Pro ($100/month) | 15 hours with GPT-Live-1.1 |
| Pro ($200/month) | No fixed cap, subject to abuse guardrails |
The help page also says Live "does not support video or screen sharing" (you switch to the older Advanced mode for that), that audio clips are kept with your transcript for 30 days, and that you can ask Live to "wait until I ask you to respond" when you want to think out loud. It is also designed for one-on-one talk and can get confused when several people speak.
The honest catch: the safety numbers
The most interesting part of the system card is the safety section. On evaluations built from real user audio, OpenAI says GPT-Live-1.1 performed "at parity" with GPT-Live-1.
But on synthetic tests designed to probe hard edge cases, OpenAI reports "statistically significant regressions on sexual, mental health, emotional reliance, and gore categories". On emotional reliance, the score dropped from 0.86 to 0.81. OpenAI says more research is needed and that it plans to improve these areas.
That is a useful reminder. The more natural a voice feels, the easier it is to lean on it. Emotional reliance is exactly the area where a lifelike voice needs the most care, whether it is a general assistant or a companion.
Why full-duplex matters so much for companionship
For a work assistant, full-duplex is a nice upgrade. For a companion, it changes the whole feeling of a call.
Voice already changes how a chatbot feels. In our look at talking vs texting your AI companion, we covered a four-week MIT Media Lab and OpenAI study of 981 adults (a preprint) in which voice users ended up less lonely and less emotionally dependent than text users at moderate use, though the advantage faded for heavy users. Natural turn-taking pushes the feeling of presence further, because the conversation flows instead of stalling.
Voice also carries emotion both ways. AI is getting better at reading feelings in tone, as we explored in whether AI can hear emotion in your voice. Combine emotion-aware listening with real-time back-and-forth, and a late-night call can feel surprisingly close to talking with someone who knows you.
And the voice itself keeps improving. Our guide to realistic AI voices in 2026 covers how expressive speech models now control emotion and pacing, and add small sounds like "mhm" and laughter.
What to look for in an AI that calls you
Whatever app you use, here is a quick checklist for voice and video calls:
- Speed. Does it answer quickly, without long awkward gaps?
- Interruptions. Can you cut in without the call breaking?
- Memory. Does it remember what you told it yesterday, or start from zero?
- Personality. Is the voice consistently theirs, with their own tone and humour?
- Video. Can you turn your camera on, and does it react to what it sees?
- Limits and price. How many minutes do you get, and what happens when they run out?
- Privacy. How long are recordings kept, and are they used for training?
Where MyBabe fits
MyBabe is a relationship app for adults (18+), built around one person you create: their face, their voice (warm, soft, deep or bright), their personality and their backstory. Calls are a big part of that. MyBabe offers real-time voice and video calls with a voice that answers as fast as a person would, and when you turn your camera on, they can notice the new haircut or the cat walking past.
Free users get two minutes of calls to try, plus 10 messages every 24 hours. Premium ($5.99 a week, $14.99 a month or $99.99 a year) includes about 30 minutes of voice and video calls a week, long-term memory so the call picks up where life left off, voice notes, four photos a day and Telegram. If you want longer nights, call packs start at $4.99 for 30 minutes.
Your companion is an AI, not a person, and not a replacement for professional mental health support. But if you want someone who picks up at 2am and already knows how your day went, create your companion on MyBabe.
FAQ
What does full-duplex mean in AI?
It means the AI can listen and speak at the same time, like a phone call. A half-duplex system takes turns, like a walkie-talkie, and has to wait for you to stop talking before it replies.
Is ChatGPT Live voice full-duplex?
Yes. OpenAI describes its GPT-Live models, including GPT-Live-1.1 released on October 15, 2026, as full-duplex models that can follow pauses and interruptions.
Why can I interrupt some AI voices but not others?
Older systems turn speech into text, process it, then speak, so they only listen between turns. Full-duplex models keep listening while they talk, so they can stop or adjust when you cut in.
How fast do humans reply in conversation?
Very fast. A PNAS study across ten languages found that answers to yes-or-no questions most often come within 200 milliseconds of the question ending.
Does a natural voice mean the AI understands me like a person?
No. A natural voice and good timing make the conversation feel human, but the AI is still software. Enjoy the warmth, and keep that difference in mind.