OpenAI published a technical blog a few days ago. In response, an author with experience developing WebRTC SFUs claims that WebRTC is not suitable for use with voice AI. [The Verge]
Experts Question OpenAI's Use of WebRTC, Warning Packet Loss Degrades Voice AI Quality
This article is a translation. Read the Japanese original
WebRTC is designed for conferencing and aggressively drops audio to minimize latency when network conditions are poor. However, the author points out that audio sent as a prompt to an LLM should prioritize accuracy, even if it means waiting approximately 200ms. Since retransmission is difficult, the author argues that degraded prompts lead to low-quality responses.
Furthermore, the author states that there are issues with the streaming of generated audio. Because WebRTC lacks a buffer and plays audio based on arrival time, OpenAI is forced to add artificial delays (sleep) before transmission to adjust the timing. [The Verge]
This method carries the risk of packets being lost and not retransmitted during network congestion. The author analyzes that this creates a result that contradicts WebRTC's objective of "maintaining low latency." [The Verge]
Source: OpenAI’s WebRTC problem (HN 511pt, 149 comments) (HN Search (backfill), 2026-05-08)