Speech-Based Dementia Detection

The field studies how changes in word choice, sentence structure, pauses, fluency, and vocal acoustics can signal dementia or milder cognitive impairment. Methods range from interpretable linguistic features and acoustic models to automatic speech recognition, self-supervised speech representations, large language models, and systems that combine audio with transcripts or other clinical data. Researchers evaluate these approaches on conversational and task-based speech benchmarks, including ADReSS, DementiaBank, and newer multilingual or clinical corpora. Strong benchmark results remain difficult to carry across datasets, languages, and patient groups, so fairness, clinical validation, faithful explanations, and robustness to recording differences are central open challenges.