Classification of silent speech words considering walking using VMD applied facial EMG

Abstract

In recent years, there has been growing interest in speech recognition technology in Japan. Against this background, several objects make use of speech recognition. However, some problems are associated with their use. For example, there is a noise problem, wherein loud ambient noise prevents accurate speech recognition. In addition, some problems depend on the user, such as the inability of people with speech impediments to use the system. Currently, there is a method for recognizing speech without speaking by attaching a myopotential sensor to the face surface. However, this method can only be recognized in a stationary state, and is difficult to recognize while walking, which involves body movements. Therefore, in this study, we focus on the recognition of silent speech while walking, which is considered difficult using conventional methods, and propose a method to improve the accuracy while minimizing the difference in classification accuracy between the case of measurement in the stationary state and the case of measurement in the walking state. In the proposed method, the variational mode decomposition algorithm, which can adaptively decompose signals, is used to finely decompose the measured signals and extract only those signals less affected by the body motion. Subsequently, based on these signals, we evaluated the proposed method using a 1-Dimensional Convolutional Neural Network (1DCNN) model that was pretrained with the measured data only at rest. Using the proposed method, we succeeded in improving the recognition accuracy of walking by approximately 10% compared to the case without it.