Fusion of descriptors for speech / music classification

This work addresses the soundtrack indexing of multimedia documents. We present a speech/music classification system based on three original features: entropy modulation, stationary segment duration and number of segments. They were merged by basic score maximisation with the classical 4 Hertz modulation energy. We validate this fusion approach with the use of the probability theory and the evidence theory. The system is tested on radio corpora. Systems are simple, robust and could be improved on every corpus without training or adaptation.