Timbre recognition of single notes using an ARTMAP neural network

In this paper, a model for the perception of musical instrument timbre is presented. The model uses an ARTMAP neural network to distinguish single notes played by five different instruments. The duration of each note is quite short. The recognition of timbre is based on three acoustic properties: spectral synchrony, slope of the attacks and spectral distribution. Arrays of values of the above properties are used as input patterns. By training the network with a large number of different input patterns a robust pattern recognizer for timbre identification is constructed. The choice of this specific type of neural network model provides the ability for creating timbre categories which can continuously be updated at any point of operation, while at the same time, knowledge of previously learned categories is retained.