This study, conducted in collaboration with Beth Israel Deaconess Medical Center and the University of Massachusetts Chan Medical School, introduces a novel approach for detecting Parkinson’s Disease (PD) using vocal biomarkers derived from speech foundation models analyzed in natural, conversational settings. Using a new US English dataset collected in clinical environments, the research demonstrates that combining features from the HuBERT Large ll60k model with an efficient Random Forest classifier yields outstanding accuracy—with an unweighted average recall (UAR) and an area under the curve (AUC) of 0.97—when screening for PD from unconstrained speech. This non-invasive, AI-driven method allows for effective screening using everyday conversation, offering both clinical-grade reliability and patient-centered flexibility.
The findings highlight several key advances: the effectiveness of modern speech foundation models for universal health screening, the robustness of these models without the need for additional fine-tuning, and validation that natural, open-ended speech can serve as a rich data source for distinguishing PD from healthy controls. Importantly, the study also finds that parameter-efficient models like Trillsson can deliver high performance in resource-constrained environments. Canary Speech’s research extends the frontier of vocal biomarker science and underscores its commitment to empowering early, accurate detection of neurodegenerative conditions, reducing burdens for patients, providers, and payers alike through innovative, voice-based technology.
Reference: Raymond Brueckner, Namhee Kwon, Vinod Subramanian, Nate Blaylock, Henry O’Connell, Luis A. Sierra, Simon Laganiere, Ella Lanzaro and Kara M. Smith. Detecting Parkinson’s Disease using Vocal Biomarkers based on Speech Foundation Models. Journal of Biomedical and Health Informatics (JBHI). Forthcoming.
