Being Old Doesn't Mean Acting Old: How Older Users Interact with Spoken Dialog Systems
Maria Wolters, Kallirroi Georgila, Johanna D. Moore, Sarah E. MacPherson · 2009 · ACM Transactions on Accessible Computing · doi:10.1145/1525840.1525842
Summary
This study challenges the common practice of designing voice interfaces based on assumed age-related characteristics. Using a bottom-up approach rather than top-down age comparisons, the researchers analyzed 447 appointment scheduling dialogs between 50 users (26 older, aged 52-84; 24 younger, aged 18-29) and nine simulated spoken dialog systems in a Wizard-of-Oz experiment. Statistical cluster analysis of linguistic features—including speech acts, word choices, politeness markers, and dialog length—revealed that users naturally fell into two distinct groups based on interaction style, not age. All participants completed a battery of cognitive tests measuring fluid intelligence, crystallized intelligence, working memory span, and information processing speed. The nine dialog system variants systematically manipulated two factors: the number of options presented (1, 2, or 4) and confirmation strategy (explicit, implicit, or none). This design allowed researchers to examine whether cognitive differences predicted interaction style and whether users adapted their behavior across multiple dialogs. The clustering methodology was rigorous: seven clustering algorithms were tested across 560 parameter combinations, with cluster quality assessed using Rousseeuw's Silhouette measure. The consistent emergence of two clusters across multiple feature sets and algorithms demonstrates the robustness of the findings.
Key findings
The analysis revealed two distinct interaction styles: "social" users (12 people) who treated the system like a human receptionist, and "factual" users (29 people) who adapted quickly to efficient, terse interactions. Social users produced roughly 350 words and 134 distinct speech acts across their dialogs, compared to just 99 words and 73 speech acts for factual users. Social users frequently used politeness markers ("please," "thank you," "goodbye"), provided unsolicited information, and engaged in meta-communication about the dialog itself. The critical finding: while 92% of social users were older, over a third (35%) of older users actually belonged to the factual cluster—behaving virtually indistinguishably from younger users. Neither age, gender, years of education, nor any of the four cognitive measures predicted which older users would fall into which group. This directly challenges age-based design assumptions. Social users were significantly less satisfied with the systems, rating them as less natural, harder to use, more frustrating, and less efficient. Factual older users, however, were actually *more positive* than younger factual users, rating the systems as more human-like and natural. Importantly, factual users adapted their behavior over time (reducing social speech acts and unprompted information), while social users showed minimal adaptation.
Relevance
This research fundamentally challenges age-based approaches to voice interface design. The finding that cognitive abilities do not predict interaction style—and that a substantial minority of older users interact just like younger users—argues against implementing blanket "elderly user" accommodations. Systems should adapt to observed behavior, not demographic assumptions. For practitioners designing voice interfaces, the implications are significant: social users pose real challenges for automatic speech recognition (richer vocabulary, unexpected phrases) and natural language understanding (politeness markers, meta-communication, overanswering). Mixed-initiative systems that can process social speech acts may improve satisfaction for this group. However, such accommodations could actually frustrate factual users who prefer efficient, streamlined interactions. The paper advocates for behavior-based adaptation: systems should detect interaction style from early dialog turns and adjust accordingly, rather than applying age-based profiles. Help prompts given early in dialogs can help shape user behavior toward more system-compatible patterns. This user-sensitive inclusive design approach recognizes the enormous diversity within older populations rather than treating them as a homogeneous group.
Tags: aging · voice interface · spoken dialog systems · cognitive aging · user diversity · natural language processing · interaction styles