Preprocessing Improvement in Mime Speech Recognition based on Surface Electromyogram

By extracting electrical signals on the surface of the speech-related muscles, which are, surface electromyography signals, information of the brain can be traced back to achieve silent speech recognition. Preprocessing is a crucial step for silent speech recognition and subsequent phases. In preprocessing of mime speech signals, baseline drift and active segment detection play an important role, but efforts in the direction are very limited. In this paper, the quadratic variation reduction separation baseline drift was introduced. On this basis, the three-parameter double threshold method based on autocorrelation function, band standard deviation, and short-time energy were used to detect the active segment of the surface electromyography signals, and good results were obtained. Finally, by extracting features, silent speech recognition was classified by the random forest algorithm. Experiments showed that the method described in this paper obtains 85% correctness in the data set of 9 words.