Text-independent speaker verification using predictive neural networks

This paper investigates the use of predictive neural networks for text-independent speaker verification. Predictive neural networks do not use discriminating information to model a speaker and hence, do not suffer from the same limitations as classification neural networks when applied to variable speaker populations. The tests were carried out using 38 speakers from the TIMIT database. Training was done using the SX sentences and testing was performed using the SA and SI sentences. The predictive neural networks developed used the two preceding frames, in a sequence of 12th order cepstral coefficients, to predict the next frame of coefficients. The results showed that the success of the system was highly dependent on the size of the hidden layer of the neural network. Two approaches for removing frames that may contain little discriminating information were tested. The better method proved to be the normalisation of each frame's score with respect to its score against all other speaker models. The use of an impostor cohort to normalise the final scores also improved the error rate.