OpenAI Details GPT-Live Engineering: Async Architecture and Go Rewrite Slash Latency

xiaohu · x · 2026-08-04

OpenAI released an engineering blog post detailing the construction of GPT-Live, its latest full-duplex real-time voice system. The new system eliminates traditional turn-taking, allowing the AI to listen and speak simultaneously with ultra-low latency.

Core Architecture: Async Fast/Slow Paths

Fast Path (Front-end Voice): Dedicated to casual chat and quick responses, passing audio data between the user and the voice model at extremely high speeds.
Slow Path (Background Brain): For complex logic, search, or tool use, it asynchronously invokes powerful models like GPT-5.5 in the background, feeding results back to the voice model without interrupting the conversation.

Network & Stack Overhaul

Migration to Go: Rewrote the media front-end and inference logic from Python to Go, significantly reducing audio packet delay jitter.
Custom WARP Protocol: An open-source spec optimized from WebRTC, reducing the 6 network round-trips required for voice connection down to 1 (or even a single UDP packet), achieving instant connection.

Original post →

OpenAI Details GPT-Live Engineering: Async Architecture and Go Rewrite Slash Latency

More from Infra