A Facial Electromyography Activity Detection Method in Silent Speech Recognition

Silent speech recognition (SSR) is a new application of human-computer interaction based on electromyography (EMG), which solves the limitation of acoustic signal dependence. In a low signal-to-noise ratio (SNR) environment, traditional methods cannot accurately segment the EMG active signal. This paper proposes an energy detection method based on spectral subtraction backtracking for detecting EMG active signals to assist silent speech recognition. The experiments are mainly based on energy detection. In addition to the energy detection, the spectral subtraction method is also used to improve the SNR and the accuracy of the endpoint information. Then, the experiments propose a backtracking method to make up for the deficiency of spectral subtraction. Finally, this paper adopts an end-to-end network model, which takes the pre-trained model of CNN as the front end, and the bidirectional gate recurrent unit (Bi-GRU) as the back end for classification. Experimental results show that the proposed activity detection method in this paper is more accurate than others in the low SNR environment.