Self-organizing neural networks and Grey's timbre space

The target of this research is an exploration of timbre multidimensionality, developed with the aid of self-organizing neural networks. Such networks show the interesting capability of extracting the main dimensions in a multidimensional input, and implement a learning algorithm derived from models of clustering in the human brain. The starting point is Grey's experiment, where a three-dimensional timbre space was determined by muldimen- sional scaling of subjective similarity judgement. Using Kohonen feature maps a timbre space was produced inside a three-dimensional neural network, allowing timbre mapping and clusterization. The input data derive directly from the Grey's music signals, after a preprocessing phase. The clusterization obtained from subjective judgement and by neural networks are compared. The results encourage the use of neural tools for timbre analysis, and suggest future developments in the fields of signal pre-processing and neural net fine tuiling.