And emotion ... part models comprising an encoder, which learns to produce a compact (or latent) representation of input speech encoding all properties of the training example, and a decoder ...