Using localized basis function for multi-speaker speech recognition

In this paper, a localized basis function neural network is suggested to perform a mathematical mapping to give a desired output vector in response to the input vector. This paper studied the mapping accuracy and convergence performance and discussed the key point of how to determine both the location and the number of the basis function, and the other key point of deciding the width of the receptive field of basis function. As an example, by way of speaker´s voice mapping, a specific radial basis function net is presented for multi-speaker speech recognition and the error rate reduced to 75% compared to the original model