Automatically selecting signal descriptors for SoundClassification

CUIDADO is a new project (European I.S.T. Project), which aims at providing content-based music applications (Vinet, Herrera, and Pachet 2002). Among these applications is an authoring tool for managing sample databases including search by similarity, search by textual attributes but also a system allowing automatic sound classification based on predefined taxonomies but also allowing user to define its own taxonomies. This last point raises a crucial issue concerning the design of the classifier but also the choice of the appropriate signal descriptors in order to perform the classification. This paper concentrates on the design of CUIDADO classifier and on two algorithms for automatically selecting the most appropriate signal descriptors for a given taxonomy: the discriminant analysis and the mutual information.