A text-to-speech synthesis method and system, a method of training a text-to-speech synthesis system, and a method of calculating an expressivity score

摘要

A text-to-speech synthesis method receives text 55 in a prediction neural network 21 trained on a first training dataset comprising text data 41a and audio data 41b for which an expressivity score 51 is calculated based on speech parameters (eg. fundamental frequency or average fundamental of several samples). A first sub-dataset 55-1 (from the first training dataset 41) is then selected 53 along with a second sub-dataset 55-2 whose average expressivity score of its audio data is higher than that of the audio data in the first sub-dataset (figs. 7a & 7b). The intermediate speech 25b is then compared 43 to converted speech 47.