Do Phonatory Features Display Robustness to Characterize Parkinsonian Speech Across Corpora?

Sustained vowels have been largely used to quantify vocal impairment in Parkinson's disease (PD), with most studies focusing on a single corpus. Presumably, features obtained from sustained vowels are language-independent, but how findings generalize across cohorts is unclear. This work analyzes 61 phonatory features from 5 corpora in American English, Italian, Castilian Spanish, Colombian Spanish, and German, respectively, by conducting a statistical and correlation analysis. We use robustness as a criterion in which a feature displays the same behavior across corpora. The statistical analysis showed that the features provided good separability between PD and controls in only two out of five corpora, and none of the features displayed robustness. However, experiments report significant correlations between feature values and clinical scores. These findings provide valuable insights into the acoustic corpora-based dissimilarities, which should be considered when generalizing findings.