Cross-lingual Retrieval-Augmented Classification for Dysarthria Severity Assessment

The article introduces Cross-lingual Retrieval-Augmented Classification (CRAC) for automatic dysarthria severity assessment, utilizing a novel align-retrieve-fuse pipeline to leverage speech data from different languages. By employing supervised contrastive learning to create a severity-focused embedding space and integrating top-k references via cross-attention during classification, CRAC achieves balanced accuracies of 87.3% on a Korean dataset and 86.7% on an Italian dataset, surpassing monolingual baselines by significant margins. This approach addresses the challenge of limited labeled pathological speech data, offering a promising method for practitioners dealing with dysarthria assessment across languages.