Raga identification of carnatic music using iterative clustering approach

This paper proposes a method to identify the arohana-avarohana of carnatic raga. Carnatic raga is broadly classified as melakarta (parent) and janya (child) raga. Arohana-avarohana of 10 different ragas is collected from 16 different singers. 16 audio data are collected for each raga. 11 among the 16 are used in the training phase and the remaining 5 are used for testing. The acoustic feature, MFCC which is increasingly used in music information retrieval, is used as the feature. Clustering model with 256 clusters is developed for all the 10 different ragas. To perform testing, the 5 test data of each raga are concatenated. Then, the sequence of training vectors (feature vectors) are divided into segments of L=100 feature vectors with 90 vectors overlapping between them. The minimum distance is calculated between each test vector and centroid of clusters. Average of the minimum distances for each segment is found out. The segment belongs to the model which has minimum of averages. To improve to accuracy, each of the ten ragas is sub classified as either parent or child group and testing is done under each group.