Singer identification using clustering algorithm

The main objective of this paper is to discuss the effectiveness of features and clustering algorithm to evaluate the performance of the singer identification system. The goal of singer identification is to identify the singer independent of training data. The training and testing phase are done for direct film song (vocal with background) for 10 singers. In training phase 15 film songs of a singer is taken as input. The input songs are made to undergo a set of pre-processing steps. The three stages of preprocessing are pre-emphasis, frame blocking and windowing. The Perceptual Linear Prediction (PLP) features are extracted from each frames of pre-processed signal. The singer model is developed by K-means clustering algorithm for each singer. In clustering method, the cluster centroids are obtained for cluster size of 256 and stored. One model is created for each singer by performing training and testing on the songs considered directly. Mean of minimum distances is computed for each model. Singer is classified based on selection of the model which produces minimum of average. The singer information is main factor in organizing and exploring music data. Singer identification also extends its application in music indexing and retrieval.