The awkward pause in AI conversations is a solved problem, at least for OpenAI. The company just pulled back the curtain on GPT-Live, its third-generation voice system, and the core insight is surprisingly straightforward: they fired the turn detector.
In a post on Wednesday, members of technical staff Justin Uberti and Zahan Malkani explained why previous voice AI felt sluggish. Systems were chained together—speech-to-text fed an LLM, which fed text-to-speech—but before any of that started, a tiny model had to guess when the human was done talking. Guess too early and you interrupt; guess too late and the response lags. GPT-Live replaces this guesswork with a full-duplex model that streams audio in and out simultaneously. It listens and speaks at the same time.
The engineering lift to make this work at scale was significant. Over six months, the team rebuilt the inference stack to be stateful, meaning it maintains continuous context rather than processing discrete blobs. A dedicated fast path shuttles audio between the client and the voice model, while a separate asynchronous boundary handles tool use and calls to frontier models like GPT-5.5. If a backend service stalls, the user’s audio stream doesn’t skip a beat.
This clean separation between the media path and application logic isn’t just about speed. It also creates a stable interface for developers. The model can now control your computer or coordinate agents in the ChatGPT desktop app without the core conversational loop ever degrading. The risk of a full-duplex system is that any delay in transport becomes an audible artifact—an unforgiving constraint that turn-based systems could mostly paper over. OpenAI’s bet is that the new architecture is tight enough to make the risk worth it, and that the resulting fluidity will make turn-based voice AI feel as dated as a walkie-talkie.
