Radial/elliptical Basis Function Neural Networks for Timbre Classification

This paper outlines a RBF/EBF neural network approach for automatic musical instrument classification using salient feature extraction techniques with a combination of supervised and unsupervised learning schemes. 829 monophonic sound examples (86% Siedlaczek Library [2], 14% other sources) from the string, brass, and woodwind families with a variety of performance techniques, dynamics, and pitches were used for the development of feature extraction, network initialization algorithms, and training of the neural networks resulting in approximately 71% individual instrument and 88% instrument family classification. A novel approach for automatically fine-tuning the system using the Nearest Centroid Error Clustering (NCC) method which determines a robust number of centroids is also discussed.