Training time and memory reduction algorithms for Speaker Recognition

The main limitation of Speaker Recognition comes from their demands in terms of training time and memory requirements. In this paper, a speaker recognizer is developed using sizable collection of various speakers with pitch and pitch strength as the feature matrix. We proposed Principal Factor Analysis (PFA) technique for dimensionality reduction and analyzed a set of SVM training algorithms suitable for training large corpora. In particular, we compare the performance of several algorithms in terms of training time, scalability towards large corpora. Our results show that the algorithms have different behaviour with respect to the time required to converge. Some of these algorithms not only scale linearly with the training set size, but also give their best results after just a few iterations.