OpenAI 详解 GPT-Live 实时语音构建：连续音频与并行推理

OpenAI explained how it built GPT-Live for real-time voice conversations.

It's built around continuous audio, i.e. GPT-Live listens and speaks continuously instead of waiting for a separate turn detector to decide whose turn it is.

Deeper reasoning and tool calls run in parallel, so the conversation keeps flowing while a stronger model works in the background.

OpenAI separated audio transport from application logic, so a slow tool call can delay its result without freezing the conversation.

They also rewrote the media frontend and inference logic in Go, reporting that the new system’s p95 frame delivery matches the previous system’s p50.

Long calls remain active through parallel handoffs, with a replacement model prefilled from the current conversation before traffic switches over.

The same mechanism compacts growing context without pausing speech, because the old model keeps talking until its replacement is ready.

Session startup uses WebRTC Abridged Roundtrip Protocol (WARP⁠) to cut WebRTC setup from six network round trips to one, while Instant Connect can begin from a single UDP packet.