Dysarthic Severity-Level Classification and Speaker Identification in Mono-Lingual and Cross-Lingual Mode

Dysarthria is a motor speech disease, which originates from neurological injury, impacts the intelligibility of human speech and affects its articulation precision. While many existing research focused on acoustic and articulatory features of speech, none of the existing studies have analyzed the influence of a speaker's language background, particularly the role of first (L1) and second (L2) languages. In this study, we investigate how L1 and L2 affects speech patterns in individuals with dysarthria at varying severity-levels, and compare with normal speech using a novel newly proposed dataset. Speech samples were collected from speakers (in their L1 & L2) diagnosed with dysarthria, as well as from normal speakers. Acoustic, prosodic, and phonetic parameters were analyzed to assess language-specific deviations and severity-related variations. We employed various acoustic features and deep learning models to analyze the effect of L1 and L2. The results indicate that L2 speakers with dysarthria exhibit distinct articulation and prosodic errors that are not commonly observed in L1 dysarthric speech. This suggests that cross-linguistic transfer effects, along with varying levels of L2 proficiency, may exacerbate motor speech difficulties. We obtained highest accuracy of 91.37 % (cross-lingual) for severity-level classification task, and 94.25 % (cross-lingual) for dysarthic speaker identification task. These findings underscore the importance of developing language-sensitive assessment tools for multilingual individuals with dysarthic disorder, highlighting the critical interplay between linguistic background and speech motor control.