Should a voice assistant say 'mm-hm' more or less depending on who's talking to it?

This explores whether tuning how often a voice agent makes small listening sounds ('mm-hm', 'I see', 'right') to suit each person would make it feel more natural than a single rate for everyone. No study in the collection tests per-user backchannel tuning directly. What it does have is two studies of the same group, older adults, that point in opposite directions, and the gap between them is the best argument for customization.

In one study, older adults rated a voice agent that could interrupt and be interrupted, and that produced backchannel cues, as more natural, engaging and fluent than a strict turn-taking agent Do voice agents that interrupt feel more natural to older adults?. In another, frequent backchannel from an LLM voice agent lowered older adults' ratings Does frequent backchannel from voice agents help or hurt older adults?. The explanation in that second study matters. Hearing slows with age, so rapid cues become hard to separate from the conversation and start to feel like interruptions instead of encouragement. Backchannel isn't good or bad in itself. Its value depends on whether the listener can take in the cues at the rate they arrive. That depends on individual hearing and processing speed, which a fixed setting can't account for.

The nearest evidence for adapting comes from pacing in text chat. Agents that changed their pacing to fit the conversation beat fixed-pace agents on perceived listening quality and human-likeness, and they also got users to share more about themselves Can pacing strategies make conversational agents feel more human?. That study was text-only and adapted to context, not to the person. Still, it shows that rhythm people perceive as attentive comes from responding to the moment, not from a fixed schedule. Research on how people mentally model dialogue partners also names 'communicative flexibility' as its own factor in how users judge an agent How do users mentally model dialogue agent partners?. However, it accounts for less variance than perceived competence or human-likeness. So getting backchannel right probably helps mainly because it makes the agent seem to be following along, more than because flexibility is valued for its own sake.

One engineering point: the pauses where backchannel would go are already in use. One system runs memory lookup inside the silence the voice-activity detector already waits through, adding no delay Can memory retrieval hide inside voice agent silence?. A per-user backchannel model would have to share those gaps. The cost of getting overlap wrong is real too. Clinical voice agents dropped required safety information in 88.6–100% of cases when patients interrupted them competitively Do patient interruptions break clinical voice agent safety?. Overlapping speech is fragile ground in both directions.

The takeaway: the collection suggests the useful setting is not 'how much backchannel' but 'how much can this listener absorb.' That depends on the listener's perception and changes with age. So the strongest form of customization may adapt to signs of how well someone is processing speech, not to their stated preferences. The collection raises this idea but doesn't test it.