Audio classification for radio broadcast indexing: Feature normalization and multiple classifiers decision

Abstract

This paper presents a system that detects the two basic components (speech and music) in the context of radio broadcast indexing. The originality of the approach covers three different points: a differentiated modelling based on Gaussian Mixture Model (GMM), which permits the extraction of speech and music components separately, the normalization of commonly used features and the efficient fusion of classifiers for speech classification which provides a substantial improvement in the presence of strong background music: accuracy of the indexing system goes from [69.2%,94.2%] for the best classifier to [90.25%,98.56%] for the fusion. Evaluation was performed on 12 hours of radio broadcast recorded under various noise conditions, channels and containing diverse speech and music mixtures. © Springer-Verlag Berlin Heidelberg 2004.