Singing Voice Detection in North Indian Classical Music

Singing voice detection is essential for contentbased applications such as those involving melody extraction and singer identification. This article is concerned with the accurate detection of singing voice phrases in north Indian classical vocal music. The component sound sources in such music fit into a typical framework (voice, rhythm and drone). We have used this a-priori knowledge to enhance the voice in the presence of accompaniment. A Gaussian Mixture Model (GMM) classifier is evaluated using frame-level feature vectors extracted from a representative data set. A threshold based method, applied to suitable audio features, is then used to automatically divide the audio signals into variable-length homogenous segments i.e. vocal or non-vocal. Segment-level classification decisions are made by grouping frame-level decisions over individual segments. The performance of the classifier is evaluated in terms of the classification accuracies of the vocal and non-vocal frames.