Corpus-Dependency in Speech-Based Parkinson's Disease Detection via SSL

Analysis of nine self-supervised learning speech backbones reveals that optimal representation layers are determined by source datasets rather than architecture. The study shows that discriminative signals for Parkinson's detection often rely on corpus-specific confounds rather than cross-lingual biological markers.