TOWARDS MUSICAL INSTRUMENT SEPARATION USING MULTIPLE-CAUSE NEURAL NETWORKS

J Klingseisen  and M D Plumbley y
mark.plumbley@kcl.ac.uk

Over the last few years, interest has been growing in neural network circles in the separation of indepen- dent sources, using techniques such as blind source sep- aration and independent component analysis (ICA). A related technique is the `Multiple-Cause Model' of Saund [Neural Computation, 7, 51-71, 1995]. In this technique, a neural network is trained to model the ob- served pattern as a composition of several underlying `causes', in contrast to the more traditional `winner- takes-all' neural networks which can handle only a sin- gle `cause'. In this paper, we report on experiments working towards the use of a simple multiple-cause model with constraints to separate di erent instruments and notes from audio spectral representations.