OpenAI's WebRTC Voice Stack Draws a 30 ms Objection That Its Own Engineers Never Addressed
Software / analysis
OpenAI's WebRTC Voice Stack Draws a 30 ms Objection That Its Own Engineers Never Addressed
Luke Curley says WebRTC drops audio older than 30 ms. OpenAI's engineers wrote about port ranges and routing, which is a different problem.
Luke Curley, who publishes as kixelated, wrote on October 27 that OpenAI still carries voice AI over WebRTC, a protocol he says discards audio once it is more than 30 ms old. OpenAI's own engineers describe a different difficulty: how to run WebRTC on Kubernetes without exposing large public UDP port ranges.
The two posts barely touch. That is the finding, and it matters for anyone choosing a transport for a voice agent.
What Curley claims about WebRTC and voice agents
Curley's post, "OpenAI still has a WebRTC Problem", follows an earlier one, "WebRTC is the problem". His argument is that "WebRTC is designed for humans, not voice AI agents." A conference drops late audio to keep the call live. A voice agent, he says, "should never ask the user to retransmit a prompt."
He makes four specific claims. Audio retransmission is off by default. The jitter buffer will not wait for a retransmit. WebRTC renders audio on arrival rather than by timestamp, which makes it hard for a text-to-speech agent to send audio faster than real time. And Google controls the protocol's direction: he says playout-delay for audio is still missing a decade after it was merged for video.
His alternative is MoQ, Media over QUIC, which retransmits within a set latency budget and sends timestamps. His worked example: one second of Opus audio at 64 Kbps is about 8 KB, sent as 50 small packets, or about 6 QUIC packets when retransmitted. He sets a 500 ms budget for real-time speech-to-speech and 5 seconds for a slower speech-to-text, model and text-to-speech chain.
- WebRTC treats audio as too old above30 ms
- Speech-to-speech budget500 ms
- Speech-to-text, model, text-to-speech budget5000 ms
Source: moq.dev, 'OpenAI still has a WebRTC Problem', October 27, 2026
Two more details show where he wants the design to go. MoQ handles interruptions by swapping tracks, and he suggests running moq-relay at CDN edges while models run wherever GPUs are available. He sizes the buffer at about 1 MB for 2 minutes of audio and proposes a max-latency of 10 minutes for low-priority backfill.
Curley says himself that he has not researched the state of voice AI. Some of his figures are rhetorical or approximate, and the page cites no sources for them.
What OpenAI's engineers built
InfoQ summarised OpenAI's post on May 20, 2026. Its authors are Yi Zhang and William McDonald, listed as members of technical staff. The constraints were global reach, fast connection setup and low, stable media round-trip times.
They rejected three options. Exposing a UDP port per session pushes complexity into Kubernetes, where large public port ranges are hard to manage safely. TURN-style relays put a heavier intermediary in the media path. Selective forwarding units suit multi-party calls, while OpenAI's sessions are mostly 1:1.
They chose a split. The relay accepts packets and forwards them, and is largely stateless. The transceiver owns ICE negotiation, DTLS handshakes, SRTP encryption and the session lifecycle. In the authors' words, "The best place to add complexity is in a thin routing layer."
Where the two accounts diverge
| Topic | Curley's post | OpenAI, per InfoQ |
|---|---|---|
| Main concern | Late audio is dropped | Public port exposure on Kubernetes |
| Retransmission | Off by default, a flaw | Not mentioned |
| Latency figures | 30 ms, 500 ms, 5 s | None in the summary |
| Session shape | Voice agents generally | Mostly 1:1 |
A relay that forwards packets does not, on the evidence of these two accounts, change what happens inside the jitter buffer. The summary places SRTP and session state in the transceiver and says nothing about audio retransmission. So the design can be sound for deployment and still leave Curley's complaint standing. Equally, his complaint is unproven for OpenAI's stack: no one has published a loss or latency measurement for it.
This report read InfoQ's summary because openai.com returned an HTTP 403 to the fetch. It cannot say what the original post includes beyond what InfoQ relays.
How the packet arithmetic reads
The 50-versus-6 packet comparison is the most concrete number in Curley's post. At 64 Kbps, a second of Opus audio is split into 50 small packets in WebRTC, which means each lost packet is a 20 ms hole. The same second fits in about 6 QUIC packets when retransmitted. He argues loss from congestion is "bursty, not random", so a burst can take out several of the 50 at once. The figures are his arithmetic, not a measurement of OpenAI's traffic, and the page says some of its numbers are approximate.
Questions for a team choosing a transport
Three checks follow from the two accounts. Does the stack retransmit audio, and for how long? Curley's example is up to 5 seconds. What does the session shape look like, 1:1 or many-to-many? OpenAI's design leans on 1:1, and the summary does not say how it would handle a conference. And who owns the stateful part? In OpenAI's design the transceiver holds ICE, DTLS and SRTP state, and InfoQ's summary names that as the tradeoff of the split.
What to watch
Curley points to Pipecat pull request 4629 as the MoQ support to follow. An OpenAI post with round-trip times under packet loss would settle the quality argument in either direction. Related reading on how agent vendors handle their own plumbing: the Windows 11 agent telemetry report and Google's Android build tooling work.
Sources
More in Software
- 01Superpowers Passes 322,000 GitHub Stars as a Set of 15 Prompted Skills for 16 Coding AgentsJesse Vincent's MIT-licensed framework tells agents to brainstorm, plan and test first. Its clone count is large, and its evidence for better results is one author's view.
- 02GitHub Disables 144 Picofly Repositories After a Nintendo Notice Names OneThe November 6 notice is the first on GitHub aimed at Switch modchip firmware, and it asked for forks to be taken down without listing a single one.
- 03Chrome's Gemini Nano Download: What Google's Docs Say, and the Numbers They Leave OutGoogle's page on built-in model management gives triggers, a 30-day rule and full-download updates, but no file size, no disk threshold and no consumer off switch.
- 04Zerodha's Blobasaur Stores Kite Layouts in Sharded SQLite, and Gives Up Durability to Do ItThe Rust, AGPLv3 blob store speaks the Redis protocol, shrank sampled production writes about 11 times, and has no live rebalancing or automatic failover yet.