Studies and Improvements in Automatic Classification of Musical Sound Samples

In this article we shall deal with automatic classification of sound samples and ways to improve the classification results: We describe a classification process which produces high classification success percentage (over 95% for musical instruments) and compare the results of three classification algorithms: Multidimensional Gauss, KNN and LVQ. Next, we introduce several algorithms to improve the sound database self-consistency by removing outliers: LOO, IQR and MIQR. We present our efficient process for Gradual Elimination of Descriptors using Discriminant Analysis (GDE) which improves a previous descriptor selection algorithm (Peeters and Rodet 2002). It also enables us to reduce the computation complexity and space requirements of a sound classification process according to specific accuracy needs. Moreover, it allows finding the dominant separating characteristics of the sound samples in a database according to classification taxonomy. The article ends by showing that good classification results do not necessarily mean generalized recognition of the dominant sound source characteristics, but the classifier might actually be focused on the specific attributes of the classified database. By enriching the learning database with diverse samples from other databases we obtain a more general classifier. The dominant descriptors provided by GDE are then more closely related to what is supposed to be the distinctive characteristics of the sound sources.