Timbre Classification by NN and Auditory Modeling

Short time Fourier analysis in combination with filter-bank techniques or cep-strum analysis have been used for many years in order to reduce timbre repre¬sentation complexity. Recently, in speech analysis and recognition, the intro¬duction of auditory models (Cooke et al. 1993) which explicitly consider non¬linear phenomena occurring in the perception mechanism, has given promising results especially when speech is highly degraded by noise (Cosi 1993). On the other hand, Neural Networks (NN) have already proved their classification capability in various pattern recognition tasks. For these reasons, a timbre clas¬sification system, directly starting from sound signals, was conceived in which auditory modeling and neural network techniques were combined together in order to reduce timbre multidimensionality. In particular S. Seneff's auditory modeling (Seneff, 1988) was used in the analysis stage, while a bidimensional Kohonen Self Organizing Map (SOM) was used in the classification stage.