Lab 04 of 5 — built by Dr. Raoul Pilcicki
The conversation you can see.
Every interruption, murmur, and pause becomes a visible state with rules — and a two-track trace you can read afterwards. Overlap is not an error — it is a state with rules. Interrupt the agent, murmur along, or stay silent, and watch each rule change the conversation.
Pick a scenario, start the agent, then interrupt it — by button, by keyboard, or with your voice level.
Press start and the agent begins to talk. It keeps listening while it speaks.
heard you — going onVoice level is optional. Only loudness is read — nothing is recorded, stored, or transcribed. A short burst counts as “mm-hm”, sustained speech as an interruption. Headphones help, otherwise the agent may hear itself.
Each rule is one behavior of the turn-taking machine. Switch one off and the next interaction changes immediately.
Yield on interruption
The agent stops mid-sentence, acknowledges you, and asks what you need. Off: it talks over you.
Continue on “mm-hm”
A short murmur is taken as encouragement — the agent notes it and keeps going. Off: every sound counts as an interruption.
Offer options after silence
At a decision point the agent holds the pause; after five quiet seconds it restates the options. Off: it waits without helping.
Typed replies are scanned against a small, open word list. The result shifts the voice — and the panel says exactly why.
Adapt pace to sentiment
Frustration slows the agent down; an easy tone lets it move a little faster.
Your track above, the agent below. After a run it stays — zoom in and read what actually happened.
Click any block to read its label and timing.
Most voice agents are built half-duplex: one side talks, the other waits. Human conversation is not like that — we interrupt, murmur agreement, and leave meaningful pauses, and we manage all of it in fractions of a second. Recent benchmark work on full-duplex dialogue names four abilities that decide whether an agent feels conversational: pause handling, turn-taking, backchanneling, and barge-in.
Duplex treats those four abilities as design material rather than model internals. Each one is an explicit, visible state with rules you can switch off — and switching one off is the fastest way to feel why it matters. The two-track timeline turns every run into an artifact you can inspect: where the agent yielded, where it kept talking through a murmur, where it held a pause instead of filling it. The interruption itself is drawn the way it happens: the agent's bar is cut, hard, at the moment you spoke.
The delivery panel borrows a second idea: agents that sense how a person is doing and adjust their style, as in Hume's empathic voice interface, and as scriptable through inline audio tags in current speech models. Here the sensing is deliberately primitive — a transparent word list over typed replies — because the point is the interaction contract: the agent must be able to say what it adapted and why.
The structured pause behavior — holding at a decision point, then offering options instead of pushing on — follows research on step-wise voice interaction, which found that structure lowers cognitive load compared to open-ended dictation.
The welcome tour speaks with a real rendered voice, embedded in this page — so the barge-in cut is a genuine mid-clip stop, and the yield, resume and offer lines are renders of their own. The other two scenarios and the answers to typed replies play silently as captions; a production build would render those too, on a full-duplex speech model with true incremental listening. On the rendered clips, pace adaptation applies as a playback-rate preview; pitch is a render-time setting. The microphone path reads loudness only — no recognition, no recording; the agent's waveform is synthetic. Sentiment is a word list, not a model, on purpose: every adaptation must be explainable in one line.
Full-Duplex-Bench v1.5 and the companion survey of full-duplex spoken dialogue models (arXiv, 2025) — the four-axis framing used here. · Hume, empathic voice interface — prosody as input, style as response. · StepWrite (2025) — structured, step-wise voice interaction reduces cognitive load.