PREDICTIVE MODELS FOR SPEAKER VERIFICATION

This paper outlines four novel methods for the task of speaker verific ation. The first model, a Hybrid Multi-Layer Perceptron (MLP)-Radial B asis Function (RBF) model, is an MLP predictor whose weights are then used as inputs to an RBF classifier for the verification process. The second model uses an array of linear predictors to model the true spea ker where each predictor is associated with a particular sub-unit of t he test utterance. The third, a Neural Prediction Model, consists of a n array of MLP predictors and the fourth, a Hidden Control Neural Netw ork, is a single MLP predictor with added control inputs. These contro l inputs modulate the MLP mapping and allow a single MLP to model a co mplete utterance. Each method was trained and tested on a modest datab ase and each performs well with verification rates of 100% for the fir st three models and of 90% for the Hidden Control Neural Network.