A Non-Intrusive Approach to Assessing Dysarthria Severity: Advancing Clinical Diagnosis

• DOI: 10.1145/3589335.3651449
• Publication Date: 2024-05-12T22:41:21Z
• Authors: Ganjun Liu; Xiaohui Hou; Meng Ge; Tao Zhang; Haizhou Li
• Keywords: Dysarthria

AI-driven severity assessment techniques for dysarthric disorders show promise in aiding speech-language pathologists with diagnostics and therapeutic follow-ups patients. Existing solutions generally focus on the average intelligibility hoarseness of individual speaker's speech (i.e., speaker-level classification). This potentially ignores slight variations pronunciation attributed to disorders, e.g., /t/ /d/. To address this issue, we rethink inherent differences dysarthria speech, propose a non-intrusive approach called DysarNet. Specifically, first design prosodic emphasis module based frame-level features highlight fine-grained temporal changes including content, rhythm, timing. Second, multi-scale aggregation strategy collect statistical cues articulatory information at different scales, i.e., utterance-level. By doing so, prosody are directly assist prediction network assessing from multiple views, naturally achieve speaker-independent generalization ability. Experimental results VCC 2018 TORGO datasets that our DysarNet excels severity.