Disentangling Style and Speaker Attributes for TTS Style Transfer

End-to-end neural TTS has shown improved performance in speech style transfer. However, the improvement is still limited by available training data both target styles and speakers. Additionally, degenerated observed when trained tries to transfer a from new speaker with an unknown, arbitrary style. In this paper, we propose approach seen unseen on disjoint, multi-style datasets, i. e., datasets of different are recorded, one individual multiple utterances. An inverse autoregressive flow (IAF) technique first introduced improve variational inference for learning expressive representation. A encoder network then developed discriminative embedding, which jointly rest modules. The proposed effectively six specifically-designed objectives: reconstruction loss, adversarial distortion cycle consistency classification loss. Experiments demonstrate, objectively subjectively, effectiveness tasks. our superior more robust than those four other reference systems prior art.