Generative x-vectors for text-independent speaker verification