Feature selection of mime speech recognition using surface electromyography data

Surface electromyography (sEMG) is a potential technique and resolution to information transmission and communication problems in noise surroundings as well as military environment. By capturing facial sEMG signals from several particular articulatory muscles, this study aims to find out features and patterns underlying sEMG data, thus decoding the speech information. In this paper, mime speech recognition on ten Chinese characters using random forest (RF) algorithm is accomplished and multiple patterns have been effectively recognized. Interpolation and alignment are applied to improve the recognition accuracy for sEMG signals with the same label. The sEMG signals are quite different in both fluctuation and span length, thus making data split and segmentation of raw data necessary. After data preprocessing, overlapped windowing techniques, which requires opportune window length and shifts, has been used to extract sEMG features. To explore proper method and feature combinations that can yield satisfactory performance, recognition results of various classification methods are compared. As an ensemble learning algorithm, random forest method is used to predict speech information and give recognition results. Experimental outcomes have indicated that the method of bicubic interpolation, alignment and windowing yields recognition accuracy that outperform other competing approaches. Classification results further demonstrate the combination of overlapped windowing technique and random forest algorithm is effective on mine speech recognition tasks.