Characterization of Hypokinetic Dysarthria by a CNN Based on Auditory Receptive Fields

Parkinson’s Disease (PD) is a major neurodegenerative disorder with steadily increasing incidence rates, demanding overgrowing resources from national health systems and imposing considerable burden on caregivers. Cost-effective and efficient turn-around time monitoring methods are required to facilitate regular, longitudinal, accurate clinical assessment and symptom management. Speech has proven to be an effective neuromotor biomarker, capitalizing on the capabilities of contact-free technology. This study aims to evaluate processing speech from people diagnosed with Parkinson’s Disease using Convolutional Neural Networks (CNN) towards characterizing speech articulation kinematics to explore differences between Healthy Controls (HC) and PD participants with Hypokinetic Dysarthria (HD), using Auditory Receptive Fields (ARFs) in the convolutional layers. The proposed proof of concept is based on a CNN described in detail, using an Extreme Learning Machine (ELM) at the output projection layer. This structure is evaluated on speech recordings from 6 PD and 6 HC participants. The performance of the approach is evaluated in terms of correlation and the log-likelihood ratio on the softmax output, showing the efficiency and retrieving properties of the CNN on speech auditory images, towards providing new insights on the pathophysiology of PD speech.