Inside OpenAI's Full-Duplex Architecture for GPT-Live Voice
The voice stack separates real-time audio from tool calls, model delegation and context maintenance to avoid audible stalls.
OpenAI published an engineering account of GPT-Live on August 3 that explains how its voice system keeps a conversation moving while other models and tools do deeper work. The official article describes a full-duplex architecture designed to listen and speak at the same time instead of waiting for a separate turn detector.
From turns to a live media path
Earlier voice systems treated each exchange as a discrete audio turn. GPT-Live removes the turn detector from the audio path and streams incoming speech into the voice model while sending generated speech back to the user. When a request needs deeper reasoning or a tool, the voice model can consult a frontier model such as GPT-5.5 without stopping the conversational flow.
The central design choice is separation. Audio travels between the client and the voice model on a dedicated fast path. Delegation, tool use and application work run behind an asynchronous boundary, so a slow backend operation can delay its own answer without blocking audio delivery. OpenAI says this foundation also supports computer control and coordination with agents in the ChatGPT desktop app.
Keeping the stream stable
OpenAI rewrote the media frontend and inference logic in Go after using a Python asyncio implementation. The company reports that the new system's 95th-percentile frame-delivery performance matched the former system's median. That is a vendor-reported relative comparison rather than an independent latency test, but it shows which engineering target the team prioritized: consistent frame timing, not only average speed.
WebRTC provides the transport layer and is used to cope with packet loss, clock drift and changes in a client connection. The system also keeps long conversations running through model-instance handoffs. It warms a replacement instance, prefills the session context, runs both instances in parallel and cuts over only when the replacement is ready.
Context compaction uses the same pattern. While the original instance continues the conversation, the system reduces accumulated context and prepares another instance with the revised state. It can then switch to the replacement without interrupting the media stream.
Why it matters
The architecture separates the work that must happen continuously from work that can finish later. That makes responsive voice less dependent on every search, tool call or model delegation completing immediately. It also gives application developers a clearer boundary: they can change tools and policies behind the asynchronous layer without redesigning the live audio path.
Status
Learning. The architecture and performance description come from OpenAI's engineering account. Internal confidence is medium because the design is first-party and the cited performance comparison is not independently reproduced in the fetched evidence.
Sources
Update note: Last reviewed 2026-08-05. We will revise this post if OpenAI publishes absolute latency data, external evaluations or additional protocol details.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.