Classification of myoelectric signal for sub-vocal Hindi phoneme speech recognition

摘要

Sub-vocal speech (SVS) recognition is highly desirable for silent communications among defence personnel and underwater operations. The SVS of Hindi phoneme has a great role to transform the sub-auditory signals into textual information for deaf Indians. Electromyography (EMG) has been applied to record signals of Hindi phoneme. EMG signals are picked up by placing electrodes over the neck areas below the chin of the subject. The SVS of Hindi phoneme recorded for four Hindi alphabets (Ka), (Kha), (Ga) and (Gha) for 10 healthy Indian subjects. Two types of features; Wavelet based features and Auto Regressive (AR) coefficient features were extracted for these phonemes. Analysis has been made using three classifiers namely linear classifier, quadratic classifier and Support Vector Machine (SVM). Performances of all three classifiers are also evaluated in terms of accuracies. The classification accuracies averaged on 10 subjects with SVM classifier are found to be 75.00%, 78.05%, 80.50% and 81.30 % corresponding to phoneme and respectively. Results also indicated that the wavelet based features with SVM classifier are best suited among three classifiers for accuracy of SVS Hindi phonemes discrimination. Myoelectric signals proved to have an important role for classification of sub-vocal Hindi phonemes in speech pattern recognition.