Voice AI assistants get talked over by patients constantly, but nobody actually counts how often — or what breaks when it happens.

This explores how often patients interrupt voice agents in real-world use. The short answer is that the corpus doesn't count it. None of these papers reports an interruption rate from a deployed system. What the collection has is evidence about what happens when interruptions occur, and evidence that interruption runs both ways. That second point may matter more than the number you came looking for.

The sharpest finding is about consequences, not frequency. When researchers deliberately had simulated patients talk over four LLM-based clinical voice agents, the agents failed to preserve required safety content 88.6–100% of the time. Prompting the models to hold their ground barely helped Do patient interruptions break clinical voice agent safety?. That's a stress test, not a field measurement. It does show that an interruption rate doesn't need to be high to be dangerous: if each interruption can drop a medication warning, even rare ones count. Compare this with a real deployment where a conversational AI took histories from 100 urgent-care patients with zero safety stops Can conversational AI safely take patient histories without supervision?. That study doesn't break out how often patients cut in, so you can't tell whether interruptions were rare or just handled well.

The lateral surprise is that the agent is often the one interrupting. A user study of spoken dialogue systems found that Moshi, a full-duplex model that can listen and speak at once, over-interrupts unpredictably. Neither system tested reliably signaled when it was holding or yielding the floor Do spoken dialogue systems handle turn-taking as well as humans?. Work on older adults goes further: an agent that could interrupt and be interrupted felt more natural than strict turn-taking, and its own topic-steering interruptions weren't resented Do voice agents that interrupt feel more natural to older adults?. Another study found that too many backchannels (small cues like "mm-hm") made the conversation worse, because older listeners process rapid audio cues more slowly Does frequent backchannel from voice agents help or hurt older adults?. So some interruption is a healthy part of conversation, and how welcome it is depends on who is listening and on timing How can proactive agents avoid feeling intrusive to users?.

There's also a reason real-world interruption counts are hard to trust. Speech recognition in noisy settings has 15–30% error rates Why do dialogue systems need probabilistic reasoning?. Many systems detect turn boundaries with voice activity detection (VAD), which treats a cough, a TV, or a mis-heard pause as a turn event. In a deployed log, "patient interruption" and "system misread the audio" can look identical. Newer single-model streaming architectures try to learn turn-taking directly instead of engineering it Can a single model learn when to speak and respond?. Some designs even use those silent gaps productively, for example running memory lookups inside them Can memory retrieval hide inside voice agent silence?.

The takeaway: frequency is the wrong first question. The corpus suggests three better ones. Does the agent keep critical content when it's cut off? Does it interrupt badly itself? Can it tell a real interruption from noise? If you want deployment numbers, that's a real gap in this collection.