Song and music discrimination play a significant role in multimedia applications such as genre classification and singer identification. Song and music discrimination play a significant role in multimedia applications such as genre classification and singer identification. The problem of identifying sections of singer voice and instrument signals is addressed in this paper. It must therefore be able to detect when a singer starts and stops singing. In addition, it must be efficient in all circumstances that the interpreter is a man or a woman or that he or she has a different register (soprano, alto, baritone, tenor or bass), different styles of music and independent of the number of instruments. Our approach does not assume a priori knowledge of song and music segments. We use simple and efficient threshold-based distance measurements for discrimination. Linde-Buzo-Gray vector quantization algorithm and Gaussian Mixture Models (GMMs) are used for comparison purposes. Our approach is validated on a large experimental dataset from the music genre database RWC that includes many styles (25 styles and 272 minutes of data).