Comparison of English and Chinese Speech Recognition Using High-Density Electromyography

Abstract

Speaking different languages requires different ways of pronunciation, and the muscular activities associated with phonation show different articulation styles. Therefore, clarifying the contributions of the articulatory muscles in different regions, such as the face and neck, is helpful for automatic speech recognition. However, it remains unclear how the articulatory muscles at different positions affect the classification accuracies of speech recognition across different languages. In this study, the technique of high-density surface electromyography (HD sEMG) was proposed to investigate the role of different articulatory muscles in classifying English and Chinese speaking tasks, respectively. The HD sEMG signals were recorded by 120 electrodes evenly placed on the facial and neck muscles across six subjects while they were speaking five English and Chinese daily words. Four time-domain features were extracted from sEMG recordings and used to construct a linear-discriminant-analysis classifier for speech recognition. The results showed that the classification accuracies of using neck sEMG were higher than that of using facial sEMG in both English and Chinese recognition tasks. The accuracies for Chinese speaking tasks were significantly higher than that for English when using facial sEMG only. Moreover, there was no significant difference in accuracies between the two types of languages when using neck sEMG. This study might provide useful information about the contributions of different articulatory muscles, and pave the way for automatic speech recognition across different languages for patients with dysarthria.