Robust Feature Extraction for Automatic Classification of Korean Traditional Music in Digital Library

Abstract

In this paper, we propose an automatic classification system that classifies the Korean traditional music in digital library. In contrast to previous works, this paper focuses on the following issues of music classification. Firstly, the proposed system accepts query sound and automatically classifies input query into one of the six Korean traditional music categories such as “Court music”, “Classical chamber music”, “Folk song”, “Folk music”, “Buddhist music”, and “Shamanist music”. Secondly, in order to overcome system uncertainty due to the different query patterns, a robust feature extraction method called multi-feature clustering (MFC) combined with SFS feature selection is proposed. Finally, several pattern classification algorithms such as k-NN, Gaussian, GMM and SVM are tested and compared in terms of the classification accuracy. The experimental results indicate that the proposed MFC-SFS method shows more stable and higher classification performance than the one without the MFC-SFS.