Full-duplex voice has spent two years chasing a metric that fits on a whiteboard: did the agent stop when the user started talking? That question produced a generation of systems tuned for interruption detection, barge-in handling, and turn-taking latency. It also produced a blind spot. Humans routinely do something the binary cannot encode — they keep the floor and weave the listener's contribution into the sentence already in flight.

Duplex Cue, described in an arXiv preprint posted September 11, 2026, is an attempt to name that blind spot and give it a measurement. According to the paper's summary, the evaluation separates listener intent into three categories — backchannel, collaboration, and interruption — and treats speaker behavior as a distinct axis. That separation is the methodological move that matters, because it lets a system be scored on whether it adapted, not merely on whether it stopped.

# What Exactly Does Duplex Cue Measure That Prior Benchmarks Miss?

The paper's core claim is that existing full-duplex evaluation collapses two different questions into one. The first question is what the listener is doing — a backchannel ("mm-hm"), a collaboration (supplying a missing word or a correction), or an interruption (taking the floor). The second is what the speaker does in response. Prior evaluations, per the paper's framing, mostly reward the speaker for yielding on any overlap, which conflates a listener who is helping with a listener who is seizing control.

Duplex Cue separates those axes. The listener-intent taxonomy is the part that will travel, because it maps onto a real conversational distinction that product teams already feel but cannot measure. A user who says "the one in Oslo" mid-sentence is not interrupting — they are completing. An agent that stops dead is not polite; it is failing to adapt.

The paper's summary explicitly names the contribution types as "a missing word, a correction or a clarification." That specificity is useful. It means the evaluation is not asking whether an agent can handle any overlap, but whether it can integrate a particular kind of semantic contribution without abandoning its own turn. That is a harder and more interesting target than barge-in detection.

# Why Is the Keep-Speaking-or-Stop Binary So Sticky?

Binary metrics survive because they are cheap to label and easy to demo. A turn-taking benchmark can be scored by a human annotator in seconds: did the agent stop? Did it keep going? Those labels are reliable, which is why they propagated across the full-duplex literature and into vendor marketing.

The cost is that the binary rewards a specific failure mode. An agent that stops on every overlap will score well on interruption handling and poorly on every conversation where the user is trying to help. According to the Duplex Cue paper, the binary "cannot express a third response humans use routinely" — continuing to speak while incorporating the listener's contribution. That is not a minor omission. It is the omission of the behavior that separates a voice assistant from a conversation partner.

My read: the stickiness is institutional, not technical. Teams ship against benchmarks they can point to in a launch post. "We reduced false interruptions by 40%" is a number. "We adapt in-turn" is a claim without a scoreboard — until now.

# What Would In-Turn Adaptation Actually Require From an Architecture?

To incorporate a mid-utterance correction, a system has to do several things at once: detect that the overlap is collaborative rather than interruptive, parse the semantic content of the contribution, revise the in-flight plan, and continue generating speech without a perceptible seam. Cascaded pipelines — VAD, ASR, LLM, TTS — struggle here because each stage introduces latency and because the LLM typically only sees the user's turn after the agent has finished speaking.

Joint speech-language models, which process input and output audio in a shared representation, are structurally better positioned for this. But the source material does not report which architectures Duplex Cue evaluated, so that inference is mine, not the paper's. What the paper does establish is the evaluation target. Once a target exists, architecture choices get pulled toward it.

This is the pattern that played out with tool use, then with long context, then with reasoning traces: a benchmark names a capability, and within two quarters the capability shows up in product roadmaps. Duplex Cue is early, but it is the kind of early that tends to stick.

# How Do the Competing Approaches Stack Up?

# What Are the Limits of This Evidence?

The source material available for this analysis is a paper summary, not the full text. That constrains what can be claimed. The summary does not report the number of models evaluated, the dataset size, inter-annotator agreement on the listener-intent taxonomy, or any quantitative results. It does not name the baselines. It does not state whether the evaluation is human-scored, model-scored, or both.

Those gaps matter because the taxonomy itself is the load-bearing element. If annotators cannot reliably distinguish collaboration from interruption in real time, the evaluation inherits the same ambiguity it was built to resolve. The paper's summary asserts the separation; it does not, in the material available, demonstrate the reliability of that separation.

There is also a construct-validity question. "In-turn adaptation" is a behavioral label applied to an observable outcome — did the agent incorporate the contribution and keep speaking? But incorporation can be superficial (the agent repeats the word back) or substantive (the agent revises its plan). The summary does not specify where the bar sits. Until that is public, Duplex Cue should be read as a framing contribution first and a scoring instrument second.

# Predictions

1. Google DeepMind will publish a full-duplex evaluation with a listener-intent taxonomy by Q2 2027, most likely as an update to its Gemini Live documentation or a companion paper, because the company's realtime API roadmap already emphasizes natural interruption handling.
2. At least one full-duplex voice startup will cite Duplex Cue in a funding or launch announcement within 12 months, using "in-turn adaptation" as a differentiator against cascaded competitors, regardless of whether it has independent replication.
3. The Duplex Cue taxonomy will be challenged on inter-annotator reliability within 18 months, either by a follow-up paper or by a competing evaluation that argues collaboration and interruption are not cleanly separable in real-time audio.

# Article Summary

• Duplex Cue's central contribution is conceptual: it separates listener intent (backchannel, collaboration, interruption) from speaker behavior, which prior full-duplex benchmarks conflate.
• The binary keep-speaking-or-stop metric rewards agents that yield on all overlap, penalizing the human-like behavior of incorporating a mid-utterance contribution while keeping the floor.
• Joint speech-language models are structurally better positioned for in-turn adaptation than cascaded pipelines, but the paper summary does not report which architectures were evaluated.
• The evidence base visible in the source material is a summary, not full results — no model count, dataset size, or inter-annotator agreement figures are disclosed, so the taxonomy's reliability is unverified.
• The framing is likely to outlast the specific benchmark: once listener intent and speaker behavior are separated, it becomes hard to justify scoring both with the same binary.