Speaker discrimination based on a fusion between neural and statistical classifiers

Abstract

Speaker discrimination consists in checking whether two (or more) speech segments belong to the same speaker or not. In this framework, we propose a new approach developed for the task of speaker discrimination, this approach results from the fusion between a neural network classifier (NN) and a statistical classifier, this fusion is obtained once by combining the scores of the simple classifiers weighted by some confidence coefficients and another time, by using the scores of the statistical classifier as an additional input of the Multi-Layer Perceptron (MLP), in order to optimize the NN training (Hybrid model). In one hand, we notice that the fusion has improved the results obtained by each approach alone and in the other hand we notice that the fusion using the sum of weighted scores, obtained by each classifier alone, seems to be better than the hybrid method. The experiments, done on a subset of Hub4 Broadcast News database, have shown the efficiency of that fusion in speaker discrimination, where the Equal Error Rate (EER) is about 7%, with short segments of 4 s only.