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, namely isolated word recognition. The dynamic nature of speech is considered by adding delayed connection and integration units to the network. We refer to a specific model where the layers are organised in a hirerchical manner: a first RBF hidden layer, a second sigmoidal layer and a classification layer which integrates over time the partial classifications performed by the sigmoidal layer. The training procedures for RBF networks are compared on both generalisation ability and computational costs. Our study shows that supervised learning of the centroids of the basis functions gives appreciable results at a significantly lower cost. In this paper we consider several learning procedures for Radial Basis Function (RBF) Networks applied to a problem of speech recognition, namely isolated word recognition. The dynamic nature of speech is considered by adding delayed connection and integration units to the network. We refer to a specific model where the layers are organised in a hierarchical manner: a first RBF hidden layer, a second sigmoidal layer and a classification layer which integrates over time the partial classifications performed by the sigmoidal layer. The training procedures for RBF networks are compared on both generalisation ability and computational costs. Our study shows that supervised learning of the centroids of the basis functions gives appreciable results at a significantly lower cost.