Improving the competitiveness of discriminant neural networks in speaker verification

The Artificial Neural Network (ANN) Multilayer Perceptron (MLP) has shown good performance levels as discriminant system in text-independent Speaker Verification (SV) tasks, as shown in our work presented at Eurospeech 2001. In this paper, substantial improvements with regard to that reference architecture are described. Firstly, a new heuristic method for selecting the impostors in the ANN training process is presented, eliminating the random nature of the system behaviour introduced by the traditional random selection. The use of the proposed selection method, together with an improvement in the classification stage based on a selective use of the network outputs to calculate the final sample score, and an optimisation of the MLP learning coefficient, allow an improvement of over 35% with regard to our reference system, reaching a final EER of 13% over the NIST-AHUMADA database. These promising results show that MLP as discriminant system can be competitive with respect to GMM-based SV systems.