OpenAI Details the Six-Month Engineering Overhaul Behind GPT-Live's Full-Duplex Voice
OpenAI published an engineering deep-dive explaining how it rebuilt ChatGPT’s voice infrastructure over six months to create GPT-Live, the full-duplex voice system introduced earlier this year. The post details the architectural rework behind natural, interruptible conversation, including a new open transport protocol called WARP and a fast-connect feature called Instant Connect.
Details
- Removing the turn detector bottleneck: Earlier voice AI architectures relied on a separate “turn detector” component to decide when a user finished speaking. GPT-Live’s full-duplex design lets the model listen and speak simultaneously instead, removing that bottleneck entirely
- WARP (WebRTC Abridged Roundtrip Protocol): A new open specification OpenAI developed to speed up connection setup by taking the SDP exchange off the critical path and collapsing the transport handshake
- Instant Connect: Paired with WARP, this lets a client start a live voice session with a single UDP packet, dramatically cutting the time from user intent to live media flow
- Asynchronous delegation: GPT-Live can consult frontier models like GPT-5.5 for complex reasoning or tool use in the background without interrupting the flow of conversation
- Stateful inference: Long-running conversations are supported through seamless handoffs between model instances and context compaction, so sessions can continue smoothly over extended periods
- Context: This is a technical follow-up to GPT-Live’s original launch, focused on the underlying engineering rather than new user-facing features
What happened next
- The improvements described are already powering ChatGPT’s voice mode across platforms
- OpenAI reiterated that an API version of GPT-Live is planned, which would let developers build on the same low-latency, full-duplex transport