Convolutional Neural Network applied in mime speech recognition using sEMG data

Decoding speaking intention from the cervical and facial muscle activity enables speech recognition independent of acoustic signals, thus allowing silent communication between human and computing devices. This research aims to design and optimize the convolutional network classifier for sEMG mime speech signals. The muscles involved in the vocalization and timbre modulation are found according to the anatomical map. Six-channel signals have been collected from the corresponding position using the sEMG signal acquisition device. The original signals are subjected to pre-processing, including noise reduction, active segment detection and interpolation, to form training and testing sets. Convolutional network models are applied to figure out the influence of structural parameters, such as convolution kernel size, the number of convolution kernels and the depth of network, on the recognition accuracy. Based on repeated trials, the optimal convolution network is provided, which provide above 80% accuracy rate.