OpenAI still grapples with a WebRTC problem
openai
| Source: HN | Original article
OpenAI continues to grapple with persistent WebRTC issues, prompting renewed criticism from the tech community.
OpenAI’s real‑time voice offering is once again under fire for its reliance on WebRTC. A post on the Media over QUIC blog, published on 27 October 2026, reignited a debate that first erupted after the company’s May 2026 blog entry describing a “rebuilt” WebRTC stack for low‑latency, global voice AI. The author, known online as kixelated, accused OpenAI of persisting with a protocol that drops audio after roughly 30 ms and plays streams by arrival time rather than timestamp, a flaw that forces the system to buffer for up to five seconds and replace tracks to cancel interruptions. The critique quickly went viral on Hacker News, where the article now carries dozens of comments and a score of 48 points.
OpenAI’s own WebRTC specialists, Sean DuBois and Justin Uberti, responded privately, according to the blog, but the technical shortcomings highlighted by kixelated remain unaddressed in the public domain. The issue is not merely academic: developers who rely on the OpenAI Realtime API have reported sudden connection failures as early as March 2025, and any latency or audio drop can degrade the conversational experience that the company markets as “seamless turn‑taking.” In a market where competitors are exploring alternative streaming protocols, persistent WebRTC glitches could erode confidence in OpenAI’s voice products and slow broader adoption.
The next weeks will reveal whether OpenAI will double down on WebRTC, adopt Media over QUIC (MoQ) or another low‑latency solution, and how quickly it can patch the reported connection failures. Industry observers will be watching for an official technical update, potential changes to the Realtime API documentation, and further community feedback on Hacker News and developer forums. The outcome could shape the technical foundation of real‑time AI voice services across the sector.
Sources
Back to AIPULSEN