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