Penerapan Model Codebook untuk Transkripsi Suara ke Teks dengan Ekstraksi Ciri Mel-Frequency Cepstrum Coefficients (MFCC)

Abstract

The advancement of information technology has triggered various demands in utilizing computer. One of them is to make computers able to communicate naturally with humans. This study uses MFCC as a feature extraction method and codebook as a pattern recognition method for voice-to-text transcription. The clustering technique used in this study is k-means. Data obtained from MFCC are clustered using the k-means method, and the model for classification is constructed using the codebook method. The utilized MFCC parameters are sampling frequency 11 000 Hz, time frame 23.27 ms, overlapping 39%. 300 voice data in WAV files with 5 seconds of duration each, are used as the training data and test data to determine the number of cepstral coefficients and the number of cluster that can produce the highest accurancy. The experiment is conducted by recognizing each syllable in 60 the test data with 240 training data. Simulation result shows that the maximum accurancy obtained is 98.3% at 26 cepstral coefficients and 12 clusters.