OpenAI Deleted the Turn Detector That Made Voice AI Feel Robotic

OpenAI has removed the turn detector, a small model that guessed when a caller had stopped speaking, from its newest voice system. That guess is why earlier voice assistants malfunctioned in two familiar ways: they cut people off mid-sentence, or they left an awkward silence before responding. GPT-Live, OpenAI’s third-generation voice architecture, ends the guessing by listening and speaking at once.

OpenAI engineers Justin Uberti and Zahan Malkani detailed the rebuild in a company engineering post published August 3. GPT-Live already drives ChatGPT Voice, including the desktop app’s new ability to operate a user’s computer and coordinate other agents. When a request needs heavier reasoning or a tool call, GPT-Live can consult a separate frontier model such as GPT-5.5 without breaking the flow of speech.

Microsoft’s unannounced MAI Realtime, which AI Insiders covered yesterday, is built on the same full-duplex idea. OpenAI’s is the version that ships, with public documentation and two named engineers explaining how it works.

Audio now moves along its own fast path, straight from the client to the voice model. Tool calls, delegation to other models, and general application logic sit behind an asynchronous boundary instead. A slow tool call only delays its own response. The live audio a caller hears keeps moving regardless.

The company also rewrote the software that moves audio and runs inference. The old version ran on Python’s asyncio. The new one is written in Go. That swap smoothed out frame delivery so consistently that the slowest normal case in the new system now performs like the typical case did in the old one. OpenAI reports that the new stack’s 95th-percentile latency matches the previous system’s median.

Long voice sessions also have to survive running out of context. Once a conversation outgrows the model’s context window, compacting it would normally invalidate the cached attention keys and values and force a slow reprocessing pause. OpenAI avoids that by treating compaction as a live handoff: a second model instance is warmed up, prefilled with the compacted context, and run in parallel with the first, then swapped in once ready. The same mechanism also handles routine instance turnover without interrupting the session.

Starting a plain WebRTC connection needs six network round trips before any media can flow. The protocol predates round-trip-conscious designs like QUIC, and it duplicates some of its own handshake steps. WARP, OpenAI’s WebRTC Abridged Roundtrip Protocol, collapses that down to a single round trip. It does this through backward-compatible tricks: folding the DTLS handshake into the ICE exchange instead of running it separately, and pre-negotiating the SCTP handshake so it does not need its own round trip.

Paired with a feature called Instant Connect, this cuts the wait further. Instant Connect negotiates connection parameters ahead of time and falls back cleanly if they go stale. Together, the two let a session start with a single UDP packet.

OpenAI is not keeping WARP to itself. The proposal is moving through a working group at the IETF. Two of the internet’s most widely used WebRTC libraries, libwebrtc and Pion, have already added support for it. A vendor pushing a genuine protocol change back into shared internet infrastructure is unusual enough to note on its own.

Before turning GPT-Live loose on real users, OpenAI quietly mirrored a small, growing slice of live ChatGPT Voice sessions to the new system in read-only mode. It ran inference behind the scenes while the existing Advanced Voice Mode kept serving the real conversation.

That test forced a rethink of what capacity means for a voice product. GPU throughput was the wrong metric, because a voice session stays open and keeps sending frames the entire time it is connected, so the CPU side of the stack has to scale alongside the model. One supporting component hit its limit earlier than load testing predicted, and requests backed up until latency compounded. The real question was never how many requests one GPU can process, but how many simultaneous sessions the system can keep alive without a single frame arriving late.

Sustained load exposed failures that short benchmarks missed. Sessions left open for hours surfaced memory and persistence pressure that never appeared in shorter tests. Reconnecting clients exercised the compaction and state restoration code paths directly. Ordinary disconnects revealed race conditions in the shutdown handshake that only showed up at scale.

One caveat is worth keeping in mind. This is OpenAI describing its own architecture, with latency figures the company chose to publish and no independent benchmark to check them against.

For teams building voice products on the Realtime API or a competing stack, the specific numbers are not the transferable lesson. The lesson is what to load test. Concurrent session survival under hours-long connections, not peak request throughput, is now the metric that decides whether a voice product holds up in production.

OpenAI engineers Justin Uberti and Zahan Malkani described the GPT-Live rebuild in a company engineering post published August 3, 2026.