Performance Comparison of Neural Networks and GMM for Vocal/Nonvocal segmentation for Singer Identification

摘要

Vocal and nonvocal segmentation is an important task in singing voice signal processing. Before identifying the singer it is necessary to locate the singer?s voice in a song. Maximum of the songs start with a piece of instrumental accompaniment known as ?prelude? in musical terms after which the singing voice comes into play. Therefore, it is necessary to detect the vocal region in the song in order to extract the singer?s voice characteristics and to avoid the non-vocal region which includes the instrumental accompaniment. This work thus classifies Vocal and Nonvocal region in songs using three different classifiers: Gaussian Mixture Model (GMM), Artificial Neural Network (ANN) with Feed Forward Backpropagation algorithm and Learning Vector Quantization (LVQ). Mel Frequency Cepstral Coefficient (MFCC) has been considered as the primary feature for classification. An available database MUSCONTENT is used and a newly created Database ASDB1 consisting of sixty excerpts from a wide variety of Assamese songs has been examined applying the same methods of classification. The efficacy of the classifiers has been tested and the results indicate that LVQ is a robust classifier compared to FFBP and GMM.