Exploration on Channel-interactive Features in Silent Speech Recognition

In recent years, silent speech recognition (SSR) based on surface electromyography (sEMG) has been proven practical and achieved successes in previous researches. The channel correlations hidden in sEMG data may work in SSR, which is rarely researched. This paper proposed to apply frame-based local binary pattern (LBP) in feature extraction of multichannel sEMG. Data are preprocessed after collection by human subjects to obtain clean signals. Then time domain (TD) features and BLP are both extracted as two feature sets, which are reduced in dimension by the method of Linear Discriminant Analysis (LDA). Random forest is used as the classifier to provide the recognitions. Experimental results show that TD obtain better result than LBP, which is exceeded by using TD and LBP simultaneously with a correct rate of 0.85. This proves LBP contains unique features in sEMG, indicating interactions in channels are promising in SSR.