Improving Dysarthria Assessment Through Voice Conversion in Low-Resource Settings

The development of automatic dysarthria assessment systems is often limited by the scarcity of labeled pathological speech, particularly in low-resource clinical environments. In this work, we explore the use of generative voice conversion to address this bottleneck. We propose a pipeline based on speech enhancement and voice conversion to transfer dysarthric vocal traits onto healthy utterances to generate realistic pathological speech. A total of 4,082 synthetic samples were generated and enhanced to improve quality while preserving pathological prosody. A dual-branch CNN trained under four experimental setups showed that combining real and synthetic data improved the classification accuracy from 77.52% to 98.36%, while using only enhanced synthetic data reached 97.10%. These results support the use of voice conversion to expand clinical datasets and reduce dependence on real patient recordings.