RAVE attacks the phase issue by using a second step of training. I don't completely understand it, but it uses a GAN architecture to make the outputs of a VAE sound better.