Robust Deep Speaker Recognition: Learning Latent Representation with Joint Angular Margin Loss