Understanding how older adults communicate with conversational agents is essential for developing reliable speech-based tools for screening and monitoring cognitive decline, including mild cognitive impairment, dementia, or Alzheimer’s disease. Such systems must also be grounded in socially and linguistically informed models of interaction to ensure ecological validity and user acceptance. This study examines whether communicative behaviour differs when older adults interact with a humanoid robot a human interlocutor in a structured speech data collection context for cognitive decline.
Fifteen participants first completed ten tasks with a human assistant (picture naming, picture description, sustained vowel, etc.), followed by a subset of six tasks with the humanoid robot Furhat. The design is limited by a fixed interaction order (human first, then robot), but this potential confound is considered in the analysis. The study combines quantitative and qualitative approaches to explore interactional differences. Prosodic features were analysed in two tasks approximating spontaneous speech: picture description and procedural description (tea preparation). We employed the measures of wiggliness and spaciousness to characterize f0 contours and intonational style, capturing macro-level variation and potential individual adaptation to different interlocutors (Wehrle, 2022). These analyses were complemented by Bayesian modelling and inference. Additionally, qualitative analyses of speaking behaviour, including conversational fillers, were conducted to examine how interlocutor type may influence task performance.
Results show that participants’ prosodic patterns remain largely stable across interlocutors, with minimal group-level differences between human-human and human-robot interactions. Inter-individual variability emerges as an important factor, suggesting that speaker-specific patterns may provide additional relevant insights.
These findings contribute to advancing socially and linguistically informed conversational AI by highlighting the relative stability of speech-based behavioural markers across interlocutor types. They further inform the design of inclusive, accessible systems that account for age-related communicative patterns, while underscoring the importance of controlled experimental designs in future work.