Speech-Based Neurological Assessment Using Voice Source Feature Analysis in Stroke Survivors

Speech production involves complex neural activity that coordinates vocal fold dynamics, articulation, and motor control. Stroke can disrupt these pathways, making glottal activity a potential biomarker for neurological impairments. However, clinical assessment of vocal fold function often requires costly, specialized equipment that limits accessibility and frequent routine monitoring. In this study, we employ glottal inverse filtering to extract voice source features from speech recordings, enabling a non-invasive, machine learning and DSP-based approach for post-stroke assessment. Using our dataset of 250 patients, we first identify significant relationships between glottal features and clinical measures of cognitive and functional stroke outcomes. The analysis is then further validated through classification tasks, distinguishing neurologically impaired from healthy speech (92 % F1-Score). These results establish speech-derived glottal features as reliable, interpretable indicators of stroke-related deficits, offering a scalable, AI-enhanced framework for objective and accessible neurological monitoring.