Could the labels linguists use for interruptions actually predict when voice AI will fail mid-conversation?
Do conversation analysis categories predict AI interruption robustness?
This explores whether the categories conversation analysts use to describe talk (types of interruption, repair, turn-taking, side-sequences) can tell us in advance which conversational moments will break an AI system, and the corpus has no paper that tests that predictive link directly.
This explores whether the vocabulary conversation analysts use for human talk (kinds of interruption, repair, turn-taking) could forecast where AI systems fail when someone cuts them off. No paper in the collection tests that prediction head-on. What it does have points one way: the categories pick out real weak spots, and current models aren't built to handle them.
The sharpest evidence comes from clinical voice agents. When patients made competitive interruptions (cutting in to take the floor, not just murmuring 'mm-hm'), four LLMs failed to preserve required safety content 88.6–100% of the time. How well each model recovered varied with no clear pattern, and telling the model to be defensive in its prompt barely helped Do patient interruptions break clinical voice agent safety?. That is a first sign the CA category matters: one named interruption type was enough to cause near-total failure. But one category in one domain falls short of a predictive taxonomy. Whether gentler interruptions, overlaps or self-repairs break models in different ways is still open.
Why would these categories line up with failure points at all? One answer is that the skills humans use to manage interruptions are social work, not information. Reference repair, topic hand-off and similar 'maintenance' moves keep a conversation running, and training that rewards predicting the next piece of information never teaches them Why don't language models develop conversation maintenance skills?. Related gaps show up in the corpus. Models don't pick up their users' word choices the way people naturally do Why don't conversational AI systems mirror their users' word choices?. And proactive moves that cut conversations by up to 60% in simulation are almost absent from AI datasets Could proactive dialogue make conversations dramatically more efficient?. If CA categories name exactly the behaviors training leaves out, they work less like a forecast of robustness and more like a map of missing skills.
Other work turns CA categories into design tools rather than tests. Insert-expansions are the side-sequences where someone pauses the main exchange to clarify. One paper uses them as a formal rule for when an agent should stop and ask the user instead of quietly chaining tool calls When should AI agents ask users instead of just searching?. On the architecture side, single streaming models that learn when to speak from one combined token stream treat turn-taking as learned behavior rather than a hand-built module Can a single model learn when to speak and respond?. That may be the more likely route to handling interruptions than prompting. A study of timing found that context-aware pacing made agents feel like better listeners Can pacing strategies make conversational agents feel more human?.
One thing you might not expect: older dialogue research already had a way to stay robust in messy conversation. It kept a probability spread over what the user might mean instead of locking in one reading, because speech recognition gets 15–30% of words wrong in noisy settings Why do dialogue systems need probabilistic reasoning?. An interruption is a similar kind of uncertainty: did the user just change their goal, or are they only signalling they're listening? The study that would answer this question would sort interruptions by CA type and measure failure rates for each. As far as this corpus shows, nobody has run it yet.
Sources 8 notes
Four LLMs tested on clinical voice tasks showed near-total failure (88.6–100%) preserving required content during competitive interruptions, with recovery varying inconsistently across models and defensive prompting offering minimal protection.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.
Response generation models fail to adapt vocabulary toward users' lexical choices, a phenomenon central to human rapport and clarity. Post-training via DPO on coreference-identified preferences can teach models in-context convention formation.
Simulations show proactivity—providing relevant information without being asked—cuts dialogue turns by 60% in medium-complexity domains. This behavior mirrors human conversation and Grice's maxims but is almost entirely absent from AI datasets and research benchmarks.
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
Show all 8 sources
Wan-Streamer represents language, audio, and video as one interleaved causal token stream, allowing response timing and turn management to be learned jointly within a single Transformer rather than engineered as separate modules, achieving sub-second latency.
A 50-person study found that conversational agents using dynamic pacing strategies outperformed static-pacing baselines on perceived listening quality, human-likeness, and user self-disclosure in text-based career and relationship support scenarios.
Real-world speech recognition achieves 15-30 percent error rates in noisy environments, making deterministic flowchart dialogue systems unworkable. POMDP-based systems handle this by maintaining belief distributions over user intent rather than committing to single interpretations.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Proactive Conversational Agents with Inner Thoughts
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- DiscussLLM: Teaching Large Language Models When to Speak
- Proactive Conversational Agents in the Post-ChatGPT World
- Conversational Alignment with Artificial Intelligence in Context
- The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMs
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- LLMs Get Lost In Multi-Turn Conversation