Style transplantation in neural network-based speech synthesis

Abstract

The paper proposes a novel deep neural network (DNN) architecture aimed at improving the expressiveness of text-to-speech synthesis (TTS) by learning the properties of a particular speech style from a multi-speaker, multi-style speech corpus, and transplanting it into the speech of a new speaker, whose actual speech in the target style is missing from the training corpus. In most research on this topic speech styles are identified with corresponding emotional expressions, which was the approach accepted in this research as well, and the entire process is conventionally referred to as “emotion transplantation”. The proposed architecture builds on the concept of shared hidden layer DNN architecture, which was originally used for multi-speaker modelling, principally by introducing the style code as an auxiliary input. In this way, the mapping between linguistic and acoustic features performed by the DNN was made style dependent. The results of both subjective or objective evaluation of the quality of synthesized speech as well as the quality of style reproduction show that in case the emotional speech data available for training is limited, the performance of the proposed system represents a small but clear improvement to the state of the art. The system used as a baseline reference is based on the standard approach which uses both speaker code and style code as auxiliary inputs.