Transkripsi Suara ke Teks Bahasa Indonesia Berbasis Suku Kata Menggunakan Codebook dan 2-level Dynamic Programming

Abstract

Voice to text transcription is very useful because it allows people to interact with a system more quickly. However, it is hard for a system to recognize a speech which contains of connected words. This research aims to develop a system that recognize a connected word speech. The proposed approach uses MFCC as a feature extraction, codebook as pattern recognition method, and 2-level dynamic programming as connected words recognition method. The parameters used in feature extraction using MFCC are overlap, time frame, and number of cepstral coefficients. Moreover the parameter used in K-means clustering is the number of clusters. This research uses 900 syllable’s speech data from 18 classes, and 120 connected word’s speech data which consist of 60 testing data joined from training data and 60 real testing data. The results showed that the optimum value of K was 15, with the minimum word error rate of 0.1 for the ‘Ide Anda’ words resulted from joining words of training data with the overlap value of 0.25, the time frame value of 25 ms, the number of cepstral coefficients of 13, and the number of clusters of 20.