Evidence from a field study with an LLM-powered voice diary
Shanshan Chen et al.
Sustaining children’s engagement and response quality in self-reporting diaries is a longstanding challenge, as repetitive tasks often reduce compliance and motivation. While conversational agents powered by large language models (LLMs) offer a promising solution, most designs for children emphasize semantic adaptation (e.g., phrasing, wording) and leave underexplored the sound-pattern form of language (e.g., rhyme/rhythm) that may shape children’s enjoyment. We conducted a seven-session field study with 32 children (aged 7–13) using an LLM-powered voice-based chatbot that supported two conversational styles: a prose baseline and a phonologically enriched (rhyming) style. Children selected their preferred style before each session, allowing analysis of behavioral preferences, compliance, response quality, and perceived enjoyment. Results show that younger children strongly preferred the rhyming style, whereas older children exhibited more varied preferences, and rhyming preference attenuated over sessions. Session-level rhyming choice was associated with higher completion, fewer skipped days before returning, and higher response quality. Mediation analyses were consistent with enjoyment as a candidate pathway linking style choice to response quality. We contribute a design account of conversational style as semantic scaffolding plus phonological enrichment, with implications for dosage-controlled and adaptive style management in agents for children.