Perbandingan berbagai metode ukuran jarak untuk pengenalan fonem dengan MFCC sebagai ekstraksi ciri

Abstract

This phoneme based voice recognition is included in the developing technology, speech to text that is a part of speech recognition. This study uses four distance measurement methods; euclidean distance, deviation standard euclidean distance, range euclidean distance and mahalanobis distance as pattern recognition, mel frequency cepstrum coefficient (MFCC) as extraction characteristic. The data used in this study consist of eleven words in Indonesian language, coba, fana, gajah, jaya, malu, pacu, quran, tip-x, visa, weda, and zakat. Phonemes are produced by segmentation process as much as twenty six phonemes, and then trained by distance measurement method to produce model. This study produces four models, using one data separation; 50% : 50%. Three overlaps; 75%, 50% and 25%. From the four distance measurement methods that used in phoneme recognition, deviation standard euclid distance method are the best of all methods, by using overlap 75% the accuracy is 89.23%.