OpenAI's Voice AI Still Drops Your Words Because of WebRTC
OpenAI still has a WebRTC Problem

WebRTC was built for conference calls, so it discards late audio packets to keep latency real-time—a disaster for voice AI, where a dropped word like "France" can leave the model guessing. Since LLMs are blocked on context, a transport could retransmit for seconds with no user impact. The author argues Media over QUIC does exactly that, buffering by timestamp instead of arrival, and that OpenAI has the clout to push Google to fix WebRTC.
A voice agent should never ask the user to retransmit a prompt. A voice agent should instead ask the transport to retransmit a prompt.