BibTeX

@inProceedings{kokkinakis-etal-2026-longitudinally-367171,
	title        = {Longitudinally Informed Synthetic Speech and Language Data for Neuropsychological Research – a Pilot Study},
	abstract     = {Large longitudinal health cohorts provide a valuable basis for studying how language, cognition, and aging interact over time, and for identifying early markers of cognitive decline. In this proof-of-concept study, we examine how such cohorts can also support the generation of synthetic neuropsychological speech and language data that preserve clinically meaningful patterns while reducing privacy risks. Using clinically grounded score profiles across individuals and time, large language models can generate synthetic cases that reflect important structural relationships without corresponding to real participants. Our pilot findings suggest that predictive modelling of cognitive trajectories remains the primary value of longitudinal cohorts, while synthetic datasets offer an important complementary resource for privacy-preserving research, benchmarking, and translational evaluation. For the text component, we explored controlled synthetic generation for task types commonly used in dementia assessment, including the Boston Naming Test, semantic verbal fluency, phonemic verbal fluency, and picture description*. Prompting strategies were designed to capture lexical retrieval difficulties, circumlocution, hesitation, omission, reduced informativeness, repetition, and topic drift. For the speech component, we investigated pipelines for speech representation, annotation, transcription, and synthetic speech generation, focusing on features such as pausing, reduced fluency, speech rate, hesitations, prosody, and temporal structure. Important challenges remain, as both natural and pathological speech are highly complex, and clinically relevant variation is expressed not only through lexical content but also through rhythm, timing, pause structure, voice quality, and prosody. Although the present findings should be regarded as preliminary, they demonstrate the feasibility of this approach and motivate future work on realism, utility, and domain adaptation for speech and language technologies used in cognitive screening and dementia-related assessment. Future work will further examine emerging text-tospeech models, with particular attention to improved prosody control, pause modelling, voice quality, temporal dynamics, and systematic validation in clinically realistic synthetic speech for dementia-related assessment contexts. },
	booktitle    = {The 36th Alzheimer Europe conference},
	author       = {Kokkinakis, Dimitrios and Chiriatti, Giuseppe and Lange, Herbert and Muñoz Sánchez, Ricardo},
	year         = {2026},
}