Speech/music discrimination based on discrete wavelet transform

In this paper we present an effective approach which addresses the issue of speech/music discrimination. Our architecture focuses on the matter from the scope of improving the performance of a speech recognition system by excluding the processing of information which is not speech. Multiresolution analysis is applied to the input signal while the most significant statistical features are calculated over a predefined texture size. These characteristics are then modeled using a state of the art technique for probability density function estimation, Gaussian mixture models (GMM). A classification scheme consisting of a conventional maximum likelihood decision methodology constitutes the next step of our implementation. Despite the fact that our system is based solely on wavelet signal processing, it demonstrated very good performance achieving 91.8% recognition rate.