Detecting Parkinson's Disease using Vocal Biomarkers based on Speech Foundation Models

Raymond Brueckner, Namhee Kwon, Vinod Subramanian, Nate Blaylock, Henry O'Connell, Luis A. Sierra, Simon Laganiere, Ella Lanzaro, Kara M. Smith

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Parkinson’s Disease, Vocal Biomarker, Foundation Models, Speech For Health, Health Care

Abstract:

Parkinson’s disease (PD) is the second most common progressive neuro-degenerative disease that leads to loss of motor control, including speech disorder. To discriminate PD from healthy control individuals, we propose an approach based on vocal biomarkers which are derived from pre-trained speech foundation models (SFM) in combination with efficient, shallow downstream classifiers. To validate our approach, we collected a new US English PD dataset in real-life clinical environments, including clinical diagnosis labels. Specifically, focusing on conversational, unconstrained speech, we compare the performance of a variety of off-the-shelf SFMs in combination with different
classifiers. We find that a combination of biomarkers derived from the HuBERT Large ll60k SFM and a Random Forest classifier leads to an unweighted average recall (UAR) of 0.97 and an Area-Under-the-Curve (AUC) of 0.97, which proves the validity of the proposed approach.

Track: 7. General Track

Registration Id: XGNKLX4K2NJ

Submission Number: 48