← Back to Explore

VoiceFormer: A Deep Learning Approach to Non-Invasive Screening of Speech Related Disorders

CWSF · 2026 Disease & Illness Bronze Medal

Thumbnail supplied by the source for VoiceFormer: A Deep Learning Approach to Non-Invasive Screening of Speech Related Disorders

Overview

Speech and voice changes can serve as noninvasive indicators of disease, particularly in respiratory, neurological, and psychiatric conditions, where subtle changes in phonation, tone, or articulation can reflect underlying pathology. However, voice-based clinical assessment is not typically integrated in diagnostic pathways, and is difficult to scale. This project investigates the utility of voice-derived features for diagnosis of diverse pathologies with known vocal changes. Using the Bridge2AI Voice Dataset, mel spectrogram representations were paired with MFCC and static acoustic features to train a dual-encoder multimodal deep learning model (VoiceFormer) for the diagnosis of 9 unique conditions. Model performance ranged between 80-98% AUROC for diseases, matching or outperforming previous state-of-the-art models. An interpretability analysis outlined the most important features for classification. This project shows the potential of using multimodal speech features and deep learning to enable real-time, accessible screening for diverse medical conditions.

Awards (2)

  • Bronze Medal
  • Selected for CWSF 2026

Competition history

  • CWSF 2026 Disease & Illness

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Source: ProjectBoard / Youth Science Canada

Save projects to your library

Sign in with Google to keep track of projects you find interesting, organized into folders. Browsing stays public.

Continue with Google