Automatic Vocal Segments Detection in Popular Music

Abstract

We propose a technique for the automatic vocal segments detection in an acoustical polyphonic music signal. We use a combination of several characteristics specific to singing voice as the feature and employ a Gaussian Mixture Model (GMM) classifier for vocal and non-vocal classification. We have employed a pre-processing of spectral whitening and archived a performance of 81.3% over the RWC popular music dataset.