Estudo sobre a robustez de técnicas de verificação de locutores com D-vectors

Recently, speaker verification techniques have been conceived through deep neural networks using d-vectors. In this work, we made experiments to compare different models in situations where we have data for training that were not obtained from the same source as the test data, representing a real problem, where it is necessary to choose a model, without having training data similar to the test data. Comparisons were made with the SincNet, GE2E and Triplet Loss Network models. This work also proposes the SincNet + GE2E model, whose performance is superior to the original GE2E. The SincNet, however, obtained the best performance results in the current scenarios.