Text independent speaker verification using multiple-state predictive neural networks

In this paper we propose a system which combines the use of predictive neural networks and the statistical approach in the task of text-independent speaker verification through the telephone line. The system is composed by a predictive neural network for every reference speaker, which is trained with the back-propagation algorithm and the maximum likelihood criterion, in order to obtain the highest probability when the input to the network belongs to the reference speaker. We also consider a global network trained on the whole training set whose likelihood gives a measure of the predictability of a given input with the aim to eliminate the strong dependence of the score from the particular input considered. In order to improve the performances of the proposed system we consider a three states ergodic model for each speaker, in this way we take into account of the non-stationarity of the speech signal.