Controlling Expressivity using Input Codes in Neural Network based TTS

摘要

This paper presents a study on the use of input codes in the neural network acoustic modeling for expressive TTS. Specifically, we use different kinds of input codes, augmented with the linguistic features, as the input of a BLSTM-based acoustic model, to control the expressivity of the synthesized speech. The input codes, in one-hot representation, include dialogue code, sentiment code and sentence position code. The dialogue code indicates whether the text is a dialogue or narration in an audiobook story. The sentiment code is obtained from a sentiment analysis tool, which labels each sentence as positive, negative and neutral. The sentence position code indicates the position of the sentence in the paragraph. We believe these codes are highly related to the expressiveness of the audiobook speech. Experiments on the data from the Blizzard Challenge 2017 demonstrate the effectiveness of the use of input codes in the neural network approach for expressive TTS.