Should AI voice assistants slow down and change their listening sounds based on how old you are?
How should conversational pacing and backchannel be tuned for different age groups?
This explores whether voice and chat agents should change how fast they talk, how long they pause, and how often they say small listening cues like 'mm-hmm' or 'right' depending on the user's age. The corpus speaks almost entirely to older adults, so it can't tell you how to tune agents for younger groups.
This explores whether conversational agents should adjust their timing and their 'I'm listening' signals for users of different ages. The corpus has a direct answer for older adults only. It has no studies on children, teens or younger adults, so it can't support a full comparison across age groups. Even so, the evidence it does have changes the question in a useful way. What matters is less 'how much backchannel' and more 'what kind, and when.'
Two studies of older adults appear to contradict each other. In one, frequent backchannel cues from an LLM voice agent lowered older adults' ratings. Their explanation is that hearing slows with age: the brain takes longer to separate rapid little sounds from the main speech, so the cues came across as noise rather than attention Does frequent backchannel from voice agents help or hurt older adults?. In the other, older adults rated an agent that backchanneled *and* let both sides interrupt as more natural, engaging and fluent than a polite turn-taking agent. They didn't even mind when the agent cut in to change the topic toward things they cared about Do voice agents that interrupt feel more natural to older adults?. Read together, the studies suggest older adults don't want a quieter, more deferential agent. They want fewer cues delivered at moments they can actually take in, inside a conversation that still feels lively and two-way. Turning everything down would probably make things worse.
The pacing research gives that idea something concrete to work with. One study broke active listening into five kinds of silence and response: reflective silence, facilitative silence, empathic silence, holding space, and immediate response. Each fits a different moment in a conversation, and an LLM could tell which moment it was in with 86% accuracy What pacing moves make conversational agents sound like active listeners?. Agents that switched among these moves were rated as better listeners and more human-like, and users told them more about themselves Can pacing strategies make conversational agents feel more human?. That points to a possible design for age: keep the same set of moves but shift the defaults, with more holding space and fewer rapid acknowledgments for listeners who need extra processing time. The corpus hasn't tested this. It's an inference.
There's a catch: current systems struggle to do any of this. A study of spoken dialogue systems found they rarely backchannel at all, and the full-duplex model Moshi interrupted too often and unpredictably Do spoken dialogue systems handle turn-taking as well as humans?. Single streaming models that learn turn-taking on their own are a promising fix Can a single model learn when to speak and respond?, but tuning by age would need a timing knob that these systems don't clearly expose yet. A deeper problem is that preference training (RLHF) pushes models away from small conversation-keeping moves like checking understanding Does preference optimization harm conversational understanding?. Language models don't pick up these relational habits from their training signal at all Why don't language models develop conversation maintenance skills?.
The point you might not expect: backchannel is not one dial. A review of how agents adapt to users found that matching someone's words improves efficiency and understanding, while matching their emotion and prosody (rhythm and tone of voice) builds warmth and trust. Mixing these up produces predictable failures Do different types of alignment serve different conversational goals?. A 'mm-hmm' sits on the prosodic and relational side. So for older adults, cutting backchannel might keep the conversation clear at the cost of warmth. That trade-off is real and should be designed on purpose, not discovered by accident.
Sources 9 notes
A CHI 2025 study found that frequent backchannel cues from an LLM voice agent reduced older adults' experience ratings. Age-related slowing in central auditory processing made rapid backchannel signals harder to distinguish and integrate into conversation, feeling disruptive rather than engaging.
Older adults rated a barge-in voice agent that interrupts mid-utterance and produces backchannel cues as more natural, engaging, and fluent than a turn-taking baseline. The agent's self-initiated topic-changing interruptions steered conversation toward participants' interests without causing resentment.
A qualitative analysis of ten active-listening videos identified five distinct pacing strategies (Reflective Silence, Facilitative Silence, Empathic Silence, Holding Space, Immediate Response), each triggered by different conversational contexts and operationalized into eight communicative moves. When implemented in an LLM agent, these strategies achieved 86.2% accuracy in real-time classification.
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.
A user study comparing Moshi (full-duplex E2E) and a VAD-based cascaded system found both rarely backchannel or signal floor-holding cues, while Moshi over-interrupts unpredictably. Even unified causal architectures did not produce better turn management than rule-based cascades.
Show all 9 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.
RLHF optimizes models for single-turn helpfulness by rewarding confident responses over clarifying questions and understanding checks. This preference alignment systematically reduces grounding acts by 77.5% below human levels, creating an alignment tax where models appear helpful but fail silently in multi-turn contexts.
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.
A 2020–2025 systematic review shows lexical alignment drives task efficiency and comprehension, while emotional and prosodic alignment drive relational warmth and trust. Conflating them in design produces category errors—cold customer-service bots and evasive mental-health assistants.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Talking Turns: Benchmarking Audio Foundation Models on Turn-Taking Dynamics
- Toward Enabling Natural Conversation with Older Adults via the Design of LLM-Powered Voice Agents that Support Interruptions and Backchannels
- Proactive Conversational Agents with Inner Thoughts
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Turn-taking and Backchannel Prediction with Acoustic and Large Language Model Fusion
- When Patients Cut In: Extending Clinical Conversational AI Safety to Interruptions
- Hear You in Silence: Designing for Active Listening in Human Interaction with Conversational Agents Using Context-Aware Pacing
- Realtime-Venus: A full-duplex interaction system with asynchronous delegation