Full-Duplex Voice

Yesterday, OpenAI announced GPT-Live-1 and GPT-Live-1 mini, a new generation of voice models that replace the old Advanced Voice Mode in ChatGPT.[1] The key shift is full-duplex: the model can speak and listen at the same time, meaning you can interrupt it mid-sentence without it talking over you, and it can pick up on conversational cues naturally.

The previous architecture was a three-stage pipeline: speech-to-text transcription, a large language model to generate a text response, then text-to-speech to deliver it. That approach worked but felt stilted. Latency stacked up across the stages. Interruptions were awkward because the system had to finish processing the current turn before it could register that you had started talking.

GPT-Live handles audio directly. It can stay silent for long stretches, absorbing context, and respond when called upon. During a press briefing, ChatGPT Voice's product lead mentioned having 30- to 40-minute conversations with the feature during walks.[2] That is a different usage pattern than the quick question-and-answer sessions most people associate with voice assistants.

The interrupt problem

Anyone who has used a voice assistant has experienced the interrupt problem. You ask a question, the assistant starts answering, you realize you want to clarify something, and you start speaking. The assistant either ignores you completely and finishes its sentence, or it stops, waits for you to finish, and then restarts from the beginning. Neither response feels like a conversation.

Full-duplex audio processing solves this at the architectural level. The model is always listening, even while producing speech. When it detects that the user has started talking, it can yield, acknowledge the interruption, and adjust. This is how human conversation works: we talk over each other, course-correct, and build on partial thoughts. Voice assistants have never been able to do that convincingly.

Why this matters for agents

The bigger claim from OpenAI is that voice could become the primary interface for complex agentic work. Instead of typing instructions to an agent that executes tasks, you talk to it, hands-free, while doing other things. The model can pull in text-based reasoning from GPT-5.5 for search, agentic capabilities, or visual formatting while maintaining the voice conversation.[2]

This resonates with work we have done on our own voice pipeline. Getting speech-to-text to work reliably with conversational audio, handling interruptions, and managing the gap between transcription and response generation are all problems we have wrestled with firsthand. The difference is that OpenAI is doing it end-to-end in a single model, while our setup chains separate STT, LLM, and TTS components. There are tradeoffs either way. A pipeline gives you modularity and the ability to swap individual components. An end-to-end model gives you lower latency and more natural conversational flow, but you are locked into whatever the model decides to do.

The competition is not standing still

Apple updated Siri in the latest iOS 27 beta to allow customization of pace and expressivity.[2] Amazon made its upgraded Alexa assistant available to everyone in the U.S. earlier this year.[2] Sesame, a startup founded by an Oculus co-founder, launched an iOS app with conversational AI that completes background tasks.[2] The race to make voice the default computing interface is crowded.

OpenAI noted that more than 150 million people already talk to ChatGPT using Voice and Dictation features.[2] That is a substantial installed base. If GPT-Live delivers on the promise of natural, interruptible, long-form conversation, it could accelerate the shift from typing to talking as the default way people interact with AI.

The demo was not perfect. A live translation showcase in Hindi had a heavy American accent and an unnatural, bookish tone.[2] The company acknowledged the model is optimized for English first. But the direction is clear: voice is no longer an add-on. It is becoming the interface.