Training-Free Cross-Lingual Dysarthria Severity Assessment via Phonological Subspace Analysis in Self-Supervised Speech Representations

This project tackles a critical 'cold start' problem in clinical speech processing: the lack of labeled pathological data for non-English languages. By leveraging frozen SSL representations (HuBERT) and projecting them into phonological subspaces defined by healthy speech, it creates a diagnostic metric that doesn't require training on dysarthric samples. Its defensibility is currently low (4) because, while the technique is academically clever, it is a standalone algorithmic approach rather than a platform. The 0 stars and 3 forks reflect its nascent status as a fresh research release (6 days old). Frontier risk is medium: while OpenAI and Google focus on general-purpose speech (Whisper, USM), they have active 'AI for Social Good' initiatives (like Google's Project Euphonia) that could eventually incorporate these specific analytical metrics. The primary moat is the specific domain expertise in phonology and subspace geometry, which is more specialized than standard fine-tuning. However, it is highly susceptible to displacement by newer, more robust SSL models (e.g., Whisper-v3 or SeamlessM4T) that might inherently capture these features more effectively.