Assessing Speech Intelligibility and Severity Level in Parkinson's Disease Using Wav2Vec 2.0

Parkinson's disease (PD) is characterized by profound speech and intelligibility impairments. This paper investigates the potential of Wav2Vec 2.0, a pre-trained speech transformer-based model, in assessing speech intelligibility and severity levels in PD. By leveraging Wav2Vec 2.0 cross-language capabilities, we deployed an English model on Italian speech data and evaluated Character Error Rate (CER). Our dataset comprised Young Healthy Controls (YHC), Elderly Healthy Controls (EHC), and PD subjects. A significant difference in the mean CER (non-parametric ANOVA; p < 0.001) was observed, with YHC being significantly different from EHC and PD. Our analysis revealed that intelligibility in the PD group did not correlate significantly with Unified Parkinson's Disease Rating Scale (UPDRS) scores (Spearman's rho = 0.37, p = 0.07). Through Z-score based detection, we were able to identify the most affected PD subjects based on their intelligibility and ranked the words that were incorrectly recognized for these individuals.