Sean DuBois is a member of the technical staff at OpenAI and the creator and longtime maintainer of Pion, the open-source Go implementation of WebRTC used in the infrastructure behind ChatGPT voice and the OpenAI Realtime API. His work brings low-latency conversation to browsers, broadcasting software, telephones, and embedded devices.
From browser video to production voice AI
DuBois discovered WebRTC while working at Etsy and continued exploring browser-based broadcasting at Infiniscene. He founded Pion WebRTC with early collaborators to give developers a more understandable, adaptable alternative to browser-centered implementations. Its modular ecosystem provides building blocks for network traversal, encrypted transport, and real-time media; he also co-authored WebRTC for the Curious, an open guide to the protocol.
At Amazon from 2018 to 2020, DuBois built a C implementation of WebRTC for embedded devices. At Apple from 2020 to 2021, he worked on browser-accessible FaceTime links and media quality-of-service systems. At Twitch from 2021 to 2024, he developed Pion-based co-streaming and WebRTC support for OBS Studio. He subsequently worked on SIP and telephony integration at LiveKit before joining OpenAI in 2024.
At OpenAI, his projects have included WebRTC server infrastructure, Realtime API support for WebRTC, SIP, and microcontrollers, 1-800-CHATGPT, and browser-control functionality for ChatGPT Operator. The company’s low-latency voice infrastructure uses a Go service built on Pion.
- Voice-to-voice latency: Conversational responsiveness depends on the complete path from microphone capture through network transport, inference, speech generation, and playback. Packet loss, jitter, buffering, and changing bandwidth are product constraints, not implementation details.
- Media-specific transport: DuBois favors WebRTC for interactive audio and video because it handles packet timing, bandwidth adaptation, and connection observability. WebSockets remain useful for prototypes, structured server-to-server traffic, and processing audio faster than real time.
- Conversational AI beyond the browser: Telephony and embedded Realtime API integrations extend voice interfaces to ordinary phones, microcontrollers, and small devices, with architectures supporting local, cloud-backed, or multiparty connections.
- Community-built real-time infrastructure: DuBois treats Pion as shared infrastructure and encourages contributors to own meaningful pieces of its development. In a public reflection on Pion’s adoption at OpenAI, he credits colleagues who scaled the system and highlights the opportunities the project created for new contributors.
