Parkinson's speech AI flops cross-lingually: layer choice varies by dataset

A new study by Serli Kopar and colleagues at arXiv (2608.13425) systematically tested whether self-supervised learning (SSL) speech models can generalize Parkinson's disease (PD) detection across languages. The team evaluated nine SSL speech backbones using a low-capacity logistic regression probe over three languages, with scenarios that progressively shifted participant identity, recording conditions, language, and pathology. The first key finding: layer selection is highly corpus-dependent. The optimal representation layer is determined primarily by the source dataset rather than the SSL architecture itself, meaning researchers cannot assume a universal "best" layer for disease detection.

This paper's second, more troubling result reveals that the transferred discriminative signal lacks pathological specificity. When classifiers trained to detect PD were applied to a target corpus, they assigned similarly high probabilities to both PD and dementia speech. That suggests the models are exploiting dataset-specific confounds (e.g., recording conditions, speaker demographics) rather than capturing disease-related motor or cognitive features. Because most SSL backbones are pretrained exclusively on healthy speech, the models appear to latch onto spurious acoustic markers. The authors conclude these limitations must be addressed before speech-based pathology recognition models can be safely deployed in clinical settings.