Toward a Speech-Based Model of Premanifest Huntington’s Disease Progression Using Deep Neural Networks

This study, conducted by researchers from Beth Israel Deaconess Medical Center, University of California San Diego School of Medicine, Harvard Medical School, and Canary Speech, explores whether speech alone can detect the earliest stages of Huntington's disease before clinical symptoms fully emerge. Huntington's disease is a progressive neurodegenerative condition marked by motor, cognitive, and psychiatric decline, yet subtle changes can begin years before diagnosis. In this study, short speech samples from individuals with premanifest and manifest disease and healthy controls were collected on tablets. Researchers extracted lexical and prosodic features-including speech rate, pauses, articulation timing, and language complexity-and evaluated multiple machine learning approaches to determine whether vocal patterns could distinguish early disease stages.

The results demonstrate that a deep neural network analyzing features from a short scripted reading task achieved approximately 81% accuracy in differentiating premanifest Huntington's disease from healthy controls, with performance improving further when identifying prodromal (83% accuracy) or manifest disease (87% accuracy). Features associated with reduced fluency such as slower speech rate, longer pauses, and lower lexical accuracy were most predictive of early disease changes. The findings support speech analysis as a scalable, objective digital biomarker capable of detecting subtle neurocognitive decline before overt motor symptoms appear. This research advances the field of vocal biomarkers and highlights the potential for voice-based screening and monitoring tools to enable earlier intervention, more precise disease tracking, and reduced burden on patients and care teams.

Reference: Luis A. Sierra, Japleen Kaur, Namhee Kwon, Vinod Submaranian, Raymond Brueckner, Nate Blaylock, Henry O'Connell, Samuel A. Frank, Jody Corey-Bloom, Simon Laganiere, Toward a Speech-Based Model of Premanifest Huntington's Disease Progression using Deep Neural Networks, Digital Biomarkers. Jan 2026.