Channel Reduction in Speech Recognition System based on Surface Electromyography

The disadvantage of surface electromyography (sEMG) application is the number of electrodes that affects to undesirable use in practical. So, this research focuses on studying the performance of the proposed sEMG-based speech recognition in classifying 12 Thai syllables by reducing the number of electrodes. The sEMG signals were acquired from the 5 facial muscles of 7 volunteers. Each sEMG signal was filtered using a band-pass filter with a cutoff frequency 20-450 Hz. Then, feature representations were extracted including feature in time domain, frequency domain and statistical domain. Next, the spectral regression extreme learning machine (SRELM) was used to reduce the dimension of data and retain most of the relevant information. Finally, feed forward neural network (NN) was applied as a classifier. The number of electrodes was reduced and its performance was tested. Results show that the average classification accuracies of 2, 3, 4 and 5 channel electrode combination were 87.05%, 94.77%, 97.81%, and 99.01%, respectively.