How we built a realtime system for responsive voice AI in six months

OpenAI published a technical account of how it built GPT-Live, a realtime system for responsive voice AI delivered in six months. The product targets continuous voice interaction rather than the stop-and-wait style of many speech interfaces, and the write-up frames the work as an engineering build, not a feature announcement alone.

GPT-Live relies on a turnless speech model paired with a low-latency architecture. Turnless design means the model does not wait for a hard end-of-turn signal before responding; audio can flow as a continuous stream. The low-latency path is the other half of the product: keeping the round trip from speech in to speech out short enough that the exchange feels like conversation instead of a sequence of recorded replies.

For engineers and builders, the useful signal is the combination of model behavior and systems design. Continuous voice only works if the speech model can operate without rigid turn boundaries and if the serving stack keeps delay low enough that interruptions, overlaps, and quick follow-ups remain usable. Anyone shipping voice agents will care less about a single clever prompt and more about end-to-end latency, streaming audio paths, and how the model handles incomplete or overlapping speech.

In the broader market, most consumer and product voice AI still behaves like turn-based chat with a microphone: the user finishes, the system processes, then the system speaks. GPT-Live’s framing—continuous interaction, turnless speech, low latency—pushes the product toward natural conversation mechanics that incumbents and startups will be measured against once users try fluid voice as a default.

Watch how applications adopt continuous, low-latency voice as a primary interface rather than a novelty mode. The concrete test for builders is whether their stacks can support turnless streaming and tight latency budgets without collapsing reliability, cost, or control over when the system should speak, listen, or stay silent.