Comparison of Classifiers for EMG based Speech Recognition

Abstract

In this paper, we propose a performance comparison of eight classifiers for speech recognition based on EMG signals to find an optimal classifier. An experiment was divided into two parts, 11 isolated Thai words classification and five Thai tones classification. The first part, EMG signals from five positions of the facial and neck muscles were captured while ten subjects uttered 11 Thai number words in both audible and silent modes. The second part, the subjects uttered 21 Thai isolated words including five tones for each word in audible mode only. Nine EMG features selected from RES index were employed and classification results of eight classifiers were compared in classification process. The results showed that a Fisher’s least square linear discriminant (FLDA) and a linear Bayes normal (LBN) classifier yielded the best result, an average accuracy was 90.01% and 79.18%, for 11 isolated Thai word classification in the audible and the silent modes, respectively. Moreover, Logistic Linear (LOGL) classifier gave the best average accuracies, of 68.36% for five Thai tone classification.