Electromyography (EMG)-based Chinese voice command recognition

This paper proposes a new method that can recognize 19 Chinese voice commands solely based on Electromyography (EMG) captured from eight facial and neck muscles. Unlike phoneme-based methods, acoustic signals were not collected herein which made the voice recognition method proposed more simplified and practical. Our focuses were on the recognition of phrases instead of isolated words to guarantee the subject a natural speaking way. For this purpose, EMG signals were segmented by detecting the start and end point of muscle activities and features representing overall information on the phrases spoken were extracted. Many features such as time domain features, frequency domain features, auto-regression coefficients, and Mel-cepstral frequency coefficients were evaluated and optimized to obtain a satisfactory feature set. Finally, linear discriminant analysis (LDA) was applied as classifier and an accuracy of 92.21% was achieved. With the proposed strategy, 19 Chinese phrases including 14 single joint motion commands and 5 control commands, namely 6 DOFs of shoulder joint, 2 DOFs of elbow joint, 6 DOFs of wrist joint, 2 on-off control commands and 3 speed control commands could be successfully recognized and it could thus provide people with a thorough training of proximal upper extremity in the early stage of rehabilitation.