MOS labels from speech synthesis boost dysarthric speech assessment

Dysarthria, a speech disorder reducing intelligibility, lacks sufficient clinically annotated data for training automatic severity assessment systems. To address this, researchers from the team behind arXiv:2606.18645 propose using human-annotated MOS labels from the QualiSpeech corpus, originally collected for speech synthesis evaluation. Their experiments show that fine-tuning a model on this synthesis data consistently improves performance on both intelligibility and naturalness prediction. Joint training with both clinical and synthesis data yields gains primarily on naturalness. This suggests synthesis artifacts and dysarthric speech share perceptual commonalities, making synthetic evaluation data a viable augmentation source.

The approach enables scalable speech monitoring and therapy-related analysis without requiring extensive new clinical annotations. By reusing existing MOS-labeled data, researchers can significantly reduce the bottleneck of data scarcity in dysarthria assessment. This work highlights a novel cross-domain transfer from speech synthesis evaluation to clinical speech assessment. Future applications could include automated severity tracking for remote therapy, personalized treatment plans, and large-scale screening tools—all made more feasible by this practical augmentation strategy.