A simple method for predicting Clinical Scores in Huntington’s Disease by leveraging ASR's uncertainty on spontaneous speech

Recent automatic speech recognition (ASR) models struggle to correctly transcribe pathological speech but implicitly capture phonetic, syntactic, and semantic properties. Unlike state-of-the-art methods that rely on speech features based on time-consuming handcrafted speech transcription, we propose a simple and fully automated approach using ASR log-probabilities to quantify intelligibility in spontaneous speech of patients with Huntington’s disease. By linking this measure to clinical scores, we explore its potential as a scalable, lightweight biomarker for disease progression. Our findings suggest that ASR-derived uncertainty offers a novel, efficient, and noninvasive alternative for clinical assessment.