SEQUENCE RECOGNITION WITH RADIAL BASIS FUNCTION NETWORKS - EXPERIMENTS WITH SPOKEN DIGITS

In this paper we consider several learning procedures for Radial Basis Function (RBF) Networks applied to a problem of speech recognition, n amely isolated word recognition. The dynamic nature of speech is consi dered by adding delayed connection and integration units to the networ k. We refer to a specific model where the layers are organised in a hi rerchical manner: a first RBF hidden layer, a second sigmoidal layer a nd a classification layer which integrates over time the partial class ifications performed by the sigmoidal layer. The training procedures f or RBF networks are compared on both generalisation ability and comput ational costs. Our study shows that supervised learning of the centroi ds of the basis functions gives appreciable results at a significantly lower cost.