COMPARISON OF FIVE CLASSIFIERS FOR CLASSIFICATION OF SYLLABLES SOUND USING TIME-FREQUENCY FEATURES

Abstract

In a speech recognition and classification system, the step of determining the suitable and reliable classifier is essential in order to obtain optimal classification result. This paper presents Indonesian syllables sound classification by a C4.5 decision tree, a Naive Bayes classifier, a Sequential Minimal Optimization (SMO) algorithm, a Random Forest decision tree, and a Multi-Layer Perceptron (MLP) for classifying twelve classes of syllables. This research applies five different features set, those are combination features of Discrete Wavelet Transform (DWT) with statistical denoted as WS, the Renyi Entropy (RE) features, the combination of Autoregressive Power Spectral Density (AR-PSD) and Statistical denoted as PSDS, the combination of PSDS and the selected features of RE by using Correlation-Based Feature Selection (CFS) denoted as RPSDS, and the combination of DWT, RE, and AR-PSD denoted as WRPSDS. The results show that the classifier of MLP has the highest performance when it is combined with WRPSDS.

Keywords

• C4.5
• Feature extraction
• Multi-layer perceptron
• Naive bayes
• Random forest
• SMO.