Abstract

Parkinson's disease is a progressive neurodegenerative disorder affectingmotor and non-motor functions, with speech impairments among its earliestsymptoms. Speech impairments offer a valuable diagnostic opportunity, withmachine learning advances providing promising tools for timely detection. Inthis research, we propose a deep learning approach for early Parkinson'sdisease detection from speech recordings, which also highlights the vocalsegments driving predictions to enhance interpretability. This approach seeksto associate predictive speech patterns with articulatory features, providing abasis for interpreting underlying neuromuscular impairments. We evaluated ourapproach using the Italian Parkinson's Voice and Speech Database, containing831 audio recordings from 65 participants, including both healthy individualsand patients. Our approach showed competitive classification performancecompared to state-of-the-art methods, while providing enhanced interpretabilityby identifying key speech features influencing predictions.