언어재활사의 노년층 마비말장애 화자 말명료도 평가와 인공지능 음성인식 시스템의 음성인식 정확도 비교 및 예측

Abstract

This study compared speech intelligibility ratings by speech-language pathologists (SLPs)(in terms of syllable intelligibility and a 5-point scale) with the speech recognition accuracy of four AI speech recognition systems (Google Assistant, Samsung Bixby, Naver Clova, and Apple Siri) for 15 older adults (aged 65 and above) with dysarthria and 15 normal elderly speakers. Based on a total of 150 sentence speech samples, the average syllable intelligibility assessed by the SLPs was 92.1%, and the average scale intelligibility score was 2.69(±0.77). A significant positive correlation was found between the two measures (R²=0.55). Among the AI systems, Clova showed the highest recognition accuracy for dysarthric speech (90.5%) and normal speech (97.3%), while Siri had the lowest accuracy (4.2% and 8.4%, respectively). A linear mixed-effects regression analysis revealed that syllable intelligibility was significantly correlated with AI speech recognition accuracy (β=0.71, p<.001), whereas scale intelligibility was not statistically significant (p=.134). Furthermore, there was a significant interaction effect between the speaker group (normal vs. dysarthric) and the type of AI system (F(3, 559.8)=12.48, p<.001). This study suggests the potential utility of AI-based speech recognition systems as auxiliary tools in speech intelligibility assessment and highlights their potential contribution to the objectification and automation of future clinical evaluations.