Silent Speech Recognition Based on Surface Electromyography

Silent speech recognition (SSR) can decode the activities of electromyography (EMG) from surface of articulatory muscles to the according brain information. As the essentials in SSR, feature extraction and recognition are directly related to the final decoding effects, which requires careful research, especially on small datasets. This paper collected the surface EMG (sEMG) of 10 isolated Chinese words by 6 surface electrodes on face and around, and based on which sEMG features were extracted by wavelet packet transform, principal component analysis and XGBoost (an implementation of gradient boosted decision trees) respectively. Then Support Vector Machine, K-Nearest Neighbor and Random Forest were explored for recognition on the 10 words. The results indicated that the highest accuracy of 72% was achieved by the method of Random Forest based on XGBoost.