Reproduce “Training-Free Cross-Lingual Dysarthria Severity Assessment via Phonological Subspace Analysis in Self-Supervised Speech Representations” on OpenResearch

About this paper

Bernard Muller, Antonio Armando Ortiz Barrañón, LaVonne Roberts

Researchers from The Scott-Morgan Foundation and Tecnológico de Monterrey developed a training-free method for cross-lingual dysarthria severity assessment using phonological subspace analysis in self-supervised speech representations. Their approach generates a 12-dimensional, clinically interpretable phonological profile by quantifying feature degradation, achieving significant correlations with severity across five languages and enabling moderate etiology discrimination and high screening potential (e.g., AUC of 0.890 for severe dysarthria detection).

Key insights

• Articulatory impairment in dysarthria measurably degrades the linear separability and distinctiveness of phonological features encoded within high-dimensional self-supervised speech representations (like HuBERT).
• Self-supervised models trained on a single language (English HuBERT) can capture universal phonological degradation patterns relevant to dysarthria severity across multiple other languages with only healthy control speech for calibration.
• Distinct patterns of phonological feature degradation emerge for different dysarthria etiologies (e.g., CP, PD, ALS), providing clinically interpretable insights into subsystem-level impairments.

Summary by alphaXiv.