The Effects of Channel Number on Classification Performance for sEMG-based Speech Recognition

Speech recognition based on surface electromyography (sEMG) signals is an important research direction with potential applications in life, work and clinical. The number and placement of sEMG electrodes play a critical role in capturing the underlying sEMG activities and in turn, accurately classifying the speaking tasks. The aim of this work was to investigate the effect of the number of channels in speech recognition based on high-density (HD) sEMG. 8 healthy subjects were recruited to perform 11 English speech tasks with sEMG signals were detected from 120 electrodes covering almost the whole neck and face. The classification accuracy was evaluated in the context of a linear discriminant analysis (LDA) with different sets of EMG electrodes. By comparing the classification accuracy, the sequential forward search (SFS) algorithm was adopted to figure out the optimal combination of electrodes which realized the highest classification level. The results showed that smaller number of channels obtained by the SFS method could achieve the classification accuracy of 80%, and another few electrodes were needed to record detail information to achieve the classification accuracy of 85%, 90% and 95%. The numbers were rather smaller than 120. Considering the computation time and reliable accuracy, it is concluded that the SFS method might be helpful for standardizing the number and position of electrodes in speech recognition.