Automatic Detection Of Vocal Segments In Popular Songs

This paper presents a technique for the automatic classification of vocal and non-vocal regions in an acoustic musical signal. The proposed technique uses acoustic features which are suitable to distinguish vocal and non-vocal signals. We employ the Hidden Markov Model (HMM) classifier for vocal and non-vocal classification. In contrast to conventional HMM training methods which employ one model for each class, we create an HMM model space (multi-model HMMs) for segmentation with improved accuracy. In addition, we employ an automatic bootstrapping process which adapts the test song’s own models for better classification accuracy. Experimental evaluations conducted on a database of 20 popular music songs show the validity of the proposed approach.