How we built a realtime system for responsive voice AI in six months
GPT-Live moves turn detection out of the voice path by making the model listen and speak continuously.
OpenAI says the system uses a full-duplex voice model, with deeper reasoning and tool calls delegated asynchronously to frontier models such as GPT-5.5. The engineering work centered on keeping audio frames moving: Go-based media and inference paths, stateful handoffs, context compaction, and WebRTC transport changes. OpenAI also describes WARP and Instant Connect, which reduce startup latency so a session can begin with a single UDP packet when pre-negotiation holds. Production shadow testing exposed capacity, geography, lifecycle, and observability problems before users heard the new system. OpenAI's note
OpenAI says the system uses a full-duplex voice model, with deeper reasoning and tool calls delegated asynchronously to frontier models such as GPT-5.5. The engineering work centered on keeping audio frames moving: Go-based media and inference paths, stateful handoffs, context compaction, and WebRTC transport changes. OpenAI also describes WARP and Instant Connect, which reduce startup latency so a session can begin with a single UDP packet when pre-negotiation holds. Production shadow testing exposed capacity, geography, lifecycle, and observability problems before users heard the new system. OpenAI's note
score 7