Locally Recurrent Probabilistic Neural Networks with Application to Speaker Verification
Todor Dimitrov Ganchev, Dimitris K. Tasoulis, Michael N. Vrahatis, Nikos Fakotakis · 2004
To improve speaker verification performance, we extend the well- known Probabilistic Neural Networks (PNN) to Locally Recurrent Probabilistic Neural Networks (LRPNN). In contrast to PNNs that possess no feedbacks, LRPNNs incorporate internal connections to the past outputs of all recurrent neurons, which render them sensitive to the context in which events occur. Thus, LRPNNs are capable of identifying time and spatial correlations. A fast three-step method is proposed for training an LRPNN. The first two steps are identical to the training of traditional PNNs, whil e the third step is based on the Differential Evolution optimization method. The per formance of the proposed LRPNNs is compared with that of the PNNs on the task of text-independent speaker verification.