Neurocomputational model of speech recognition for pathological speech detection: a case study on Parkinson's disease speech detection

This paper presents a computational model for distinguishing between healthy speech and pathological speech, specifically speech from patients with Parkinson's disease. The model is based on neurophysiologically plausible computational models of speech and syllable recognition. These models were designed to uncover the functional roles of brain activity during speech perception. The proposed model is a two-level generative model that uses predictive coding to identify whether the input syllable corresponds to the healthy or Parkinson's disease condition. During inference, the model accumulates the evidence associated with each condition. Although early results are modest (around 60% AUC), they suggest that this approach has merit and should be further investigated.