Using self-supervised speech representations to track Parkinson’s Disease severity

Parkinson’s Disease (PD) affects motor control, including speech, yet identifying reliable acoustic markers of disease progression remains a challenge. This study investigates whether speech embeddings derived from patient recordings reflect clinical disease severity. We analyzed vowel recordings from Castilian Spanish-speaking PD patients using representations from a wav2vec 2.0 XLSR-53 model fine-tuned on the Spanish portion of the Common Voice dataset. Dynamic time warping (DTW) was applied to compare each patient vowel against a reference set of healthy controls. The resulting distances, normalized by alignment path length, were examined in relation to Hoehn and Yahr (H–Y) stage using correlation and regression analysis. DTW distance was not significantly associated with H–Y stage (ρ = 0.08, p = 0.08; p = 0.74 in regression). Vocal tremor showed a modest correlation with embedding distance (r = 0.19), but did not predict disease severity. In contrast, hypophonic voice, dysphagia, sialorrhoea, and age were all significant predictors of H–Y stage. These results suggest that mid-layer speech embeddings may not capture key vocal-motor impairments associated with disease progression. Future work will explore suprasegmental and prosodic features (e.g., intensity, shimmer), alternative embedding layers, and direct modeling of symptom-specific targets like hypophonia rather than global severity scales.