Penerapan Mel Frequency Cepstrum Coefficients (MFCC) sebagai ekstraksi ciri pada pengenalan fonem dengan Probabilistic Neural Network (PNN) sebagai classifier

Abstract

Voice recognition (speech recognition) is one field of study in voice processing. This technology can convert voice signals into a form of written information (text). With this technology, people can interact with a computer. MFCC feature extraction computes the cepstral coefficients by considering the human hearing. This research is a phoneme recognition using feature extraction MFCC and PNN as feature matching model. This study compares the coefficient, overlap, and the test data without noise and with noise. MFCC is used with 13, 20, 26 coefficients and 0%, 25%, 50% overlap. Noise is added by 30 dB, 20 dB and 10 dB. From the comparison of the three overlaps, produced the best accuracy at 50% overlap with an accuracy of 94,71%. From comparing the three coefficients, better accuracy resulting in coefficients of 20 and 26 with an accuracy of 97,12% at 50% overlap. After comparing between the coefficients of 20 and 26 with overlap of 25% then the coefficient 26 is obtained that better accuracy of 94,23%. This shows that in this study the coefficient 26 is the best. In this research, there are three noise variables. The variables are 10 dB, 20 dB, and 30 dB. The best accuracy reached when the noise variable is 30 dB rather than 10 dB or 20 dB because the accuracy has the closest accuracy compared by the accuracy when noise was not added into the data. When the noise variable is 30 dB, the percentage of accuracy is 85,3%.