Pemodelan Sistem Identifikasi Pembicara Dengan MFCC dan Support Vector Machine

Abstract

In this paper, we focus on speech recognition were speakers used text-dependent which means the text agreed in advance and will be used next. This system using MFCC as feature extraction and SVM as pattern recognition. Data were taken from 10 adult speakers with differences in gender, age and ethnicity. Each speakers provide 50 ballot "computer" and its pronunciation is not controlled resulting in 500 data. Some data training are contaminate with gaussian noise with level 80dB, 70dB, 60dB, 50dB, 40dB, 30dB, 20dB, 10dB and 0dB. The research uses a frame length: 40 ms, overlapping frames: 50%, and the coefficient mel: 13. Noise Cancelling also tested in this research, although not getting optimal results. Pattern recognition SVM with RBF kernel functions produce 100% accurate results. Time process of Sequential Minimal Optimization algorithm is better than Quadratic Programming algorithms. Increasing the number of speakers to see the performance of the system with a greater amount of data can be made for further research.