Hybrid Algorithm Applied to Feature Selection for Speaker Authentication

One of the speaker authentication problems consists on identifying a person only by means of his/her voice. To obtain the best authentication results, it is very important to select the most relevant features from the speech samples, this because we think that not all of the characteristics are relevant for the authentication process and also that many of these data might be redundant. This work presents the design and implementation of a Genetic-Neural algorithm for feature selection used on a speaker authentication task. We extract acoustic features such as Mel Frequency Cepstral Coefficients, on a database composed by 150 recorded voice samples, and a genetic feature selection system combined with a time delay feed-forward neural network trained by scaled conjugate gradient back propagation, to classify/authenticate the speaker. We also show that after the hybrid system finds the best solution, it almost never looses it, even when the search space changes. The design and implementation process, the performed experiments, as well as some results are shown.