Clustering of Instruments in Carnatic Music for Content Based Information Retrieval

Music Information Retrieval (MIR) focuses on retrieving useful information from collection of music. The objective of research work in this paper is to explore clustering approaches which can be useful in automatically mining the content from Carnatic instrumental music. The content to be retrieved is the instrument that is primarily used to play the song. Carnatic music songs with ten different instruments namely, Flute, Harmonium, Mandolin, Nagaswara, Santoor, Saxophone, Sitar, Shehnai, Veena and Violin are considered as input. Mel Frequency Cepstral Coefficients (MFCC) and Linear Predictive Coefficients (LPC) features are used for representing music information. In the first step, visualization technique is used to explore the capability of different features in distinguishing Carnatic music with different instruments. Then different clustering techniques are used for understanding natural way of grouping among this instrumental music. A discussion on the comparison of instrument clustering results with different algorithms, combined with various features is also presented.