Syllable-Level Acoustic Modeling with a Stage-Aware Transformer for ALS Dysarthria Severity Estimation - ICASSP-2026 Sand Challenge

This paper presents a syllable-level Transformer-based model for estimating dysarthria severity in patients with Amyotrophic Lateral Sclerosis (ALS) as part of the ICASSP-2026 SAND Challenge [1]. Our approach segments speech recordings into syllable-level acoustic units using Sylber and extracts eGeMAPS Low-Level Descriptors (LLDs). We then compute feature-level mutual information scores to rank and select the most discriminative features. After integrating features extracted from three distinct syllable recordings (/pa/, /ka/, and /ta/), we employ a stage-aware Transformer framework with separate models trained for each ALSFRS-R starting severity. Experimental results demonstrate that this syllable-level, stage-wise modeling improves the prediction of future ALS dysarthria severity.