This recipe replaces ivectors used in the v1 recipe with embeddings extracted from a deep neural network. In the scripts, we refer to these embeddings as "xvectors." The recipe is closely based on the following paper: http://www.danielpovey.com/files/2018_icassp_xvectors.pdf but uses a wideband rather than narrowband MFCC config. In addition to the VoxCeleb datasets used for training and evaluation (see ../README.txt) we also use the following datasets for augmentation. MUSAN http://www.openslr.org/17 RIR_NOISES http://www.openslr.org/28