OpenAI recently published a technical blog that sparks debate about using WebRTC for voice AI. The post argues the protocol is ill‑suited for conversational agents because it prioritizes real‑time latency over accuracy. The author, a former Twitch engineer, points to his experience rewriting WebRTC modules in Go and Rust across multiple production systems.

WebRTC’s aggressive packet dropping creates audible glitches when network conditions dip, a problem the author highlights after testing with Discord. Because the protocol’s jitter buffer caps at 200 ms, users experience distorted audio rather than a brief pause. For a paid voice assistant, losing prompt fidelity costs more than a few milliseconds of delay.

OpenAI’s approach involves routing only STUN packets through a single UDP:443 port, then caching DTLS, RTP, and RTCP state. This bypasses the need for multiple ports but forces the system to assume the client’s source IP never changes—a fragile assumption when users switch Wi‑Fi to cellular. The result is an 8‑round‑trip setup that defeats the promised low‑latency target.

In practice, the chosen design forces OpenAI to trade packet reliability for speed, leaving users to endure occasional audio drops. The trade‑off illustrates a broader tension in real‑time AI services: whether to lean on proven protocols like WebRTC or build custom solutions that expose new failure modes. Ultimately, the decision shapes user trust and the product’s scalability.