An Empirical Study on Text-Independent Speaker Verification based on the GE2E Method

While many researchers in the speaker recognition area have started to replace the former classical state-of-the-art methods with deep learning techniques, some of the traditional i-vector-based methods are still state-of-the-art in the context of text-independent speaker verification. Google's Generalized End-to-End Loss for Speaker Verification (GE2E), a deep learning-based technique using long short-term memory units, has recently gained a lot of attention due to its speed in convergence and generalization. In this study, we aim at further studying the GE2E method and comparing different scenarios in order to investigate all of its aspects. Various experiments including the effects of a random sampling of test and enrollment utterances, test utterance duration, and the number of enrollment utterances are discussed in this article. Furthermore, we compare the GE2E method with the baseline state-of-the-art i-vector-based methods for text-independent speaker verification and show that it outperforms them by resulting in lower error rates while being end-to-end and requiring less training time for convergence.