DEVOURED
How OpenAI Built GPT-Live
OpenAI transitioned its voice architecture to a full-duplex, stateful inference model to support simultaneous listening and speaking with low-latency feedback.
What: The new GPT-Live architecture replaces traditional request-response cycles with a stream-oriented design, utilizing asynchronous delegation and dynamic context management to maintain conversational responsiveness during complex reasoning and tool use.
Why it matters: This reflects the move toward 'agentic voice' interfaces that prioritize real-time interactivity and context-aware flow, essential for agents to participate naturally in human environments.
Deep dive
- Full-Duplex: The ability to send and receive voice data simultaneously, preventing the 'walkie-talkie' delay found in standard voice chat.
- Stateful Inference: Maintaining conversation state across voice turns to avoid repetitive prompts or context loss.
- Asynchronous Delegation: Moving background processes or tool calls off the main voice loop so the user's conversation is not interrupted while the model thinks.
Decoder
- Full-Duplex: A communications system that allows data transmission in both directions at the same time.
Original article
OpenAI rebuilt its voice architecture around a full-duplex model that listened and spoke simultaneously. The system combined stateful inference, asynchronous delegation, dynamic context management, and low-latency media transport to keep conversations responsive while supporting advanced reasoning and tool use.
Explore all 45 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi