Performance analysis of wavelet transform for unspoken words

Abstract

This paper aims to identify unspoken words using facial muscle activity without audible signal. A novel approach of wavelet analysis using different filter coefficients and the performance analysis of each filter coefficients along with classifiers is introduced. The technique is successfully used to classify 5 vowels, month name, weekdays and is not sensitive to the variation in speed and the speaking style of the different subjects and is used as recognition variables to recognize the unspoken words. This is an efficient technique that measures the relative muscle activity of the articulatory facial muscles and it is shown that former systems go through austere performance degradation when new words are acquainted. Using wavelet, word error rate and classification rate achieve a noteworthy success. This paper presents wavelet analysis of surface electromyogram (SEMG) of four facial muscles to decompose the signal and identify the words and classifying these words by three wavelets. Using three filter coefficients of Daubechies, Coiflets and Symlets, the unspoken word has been recognized. This represents the proportional muscle activity of the four muscles. These are classified using k-nearest neighbor, probabilistic neural network and support vector machine to identify the speech. It is urged that such applied science may be used for the user to give simple voiceless commands when conditioned for the specific user.