OpenAI unveils real-time two-way voice in GPT-Live-1; Yelp first to deploy it for restaurant bookings

OpenAI has revealed a new capability in GPT-Live-1: the model can listen while it speaks, letting callers interrupt, add details or change course mid-conversation without waiting for a turn. The first live implementation comes from Yelp, which is embedding the model in its phone-based restaurant reservation system.

The technical advance is a full-duplex architecture. The system processes incoming audio while simultaneously generating outgoing audio, rather than pausing until the other party finishes a sentence. OpenAI says this enables natural interruptions — changing the party size, asking for a window table, cancelling the booking mid-sentence — without the dialogue stalling in a "hold" state.

Yelp is demonstrating the capability in a table-booking scenario. A caller speaks with the voice agent, and the model reacts in real time as the request shifts. The integration is meant to handle the way reservation calls tend to drift off script — questions about the menu, allergies, seating preferences — into a flow that feels closer to talking with a human host.

OpenAI did not disclose latency figures, speech-recognition error rates in noisy conditions, API availability, pricing or conversation-length limits. It also remains unclear whether the model will be offered to developers beyond the Yelp partnership, or whether it stays in a closed demo phase.