Application of Fractal Analysis based Feature Extractor for Channel Reduction of Silent Speech Interface Using Facial Electromyography

Abstract

Surface electromyography (sEMG) based silent speech interface (SSI) is an actively investigated topic among the broad area of human computer interaction studies which is currently dominated by acoustic sound based speech recognition research. This research is an attempt to help people who have an impaired vocal system if they are having no issues with their facial muscle functions. The basic idea is to reduce the total number of sEMG electrodes that has to be affixed on the face thereby reducing the invasiveness of the silent speech recognition module. This is achieved by incorporating a new detrended fluctuation analysis (DFA) based feature along with the already existing features associated with electromyographic signals. DFA is used for the first time in literature in the area of surface electromyography based silent speech recognition. The main idea is to incorporate the DFA feature along with the state-of-the-art features to improve the performance of a sEMG based SSI model so that an efficient channel reduced model can be realised. Different channel combinations were tried to analyse the impact of each channel in word recognition accuracy and the optimal channel combination was identified. As a result of this research work, a reduced channel setup with 5 electrodes was proposed in place of the conventional 7 channel data acquisition setup. This was achieved while maintaining an accuracy of 83.88 % and 92.92 % using the decision tree (DT) model and K-nearest neighbours (KNN) model respectively.