A study on speech diversification techniques based on corpus design for advanced humanoid speech synthesis

我们的研究目标是实现具有多种情感表达和说话风格的更接近人类、更自然的文本到语音系统，我们的研究成果如下：(1)我们提出了一种新的语料库设计技术，该技术考虑了口音、风格和句子最终表达。(2)将用户主观情绪强度纳入声学模型训练，提高表达性语音合成性能。(3)提出了一种基于基频参数生成技术的强调表达自动标注技术，实现了强调语音合成。(4)提出了仅使用目标说话人的母语语音样本进行跨语言语音合成，以低成本合成多语言语音。

Our goal in this research is to realize more human-like, natural text-to-speech system with various emotional expressions and speaking styles, and the achievements of our studies are as follows:(1)We proposed a novel corpus-design technique in which accent, style, and sentence-final expression are taken into account. (2)We incorporated user's subjective emotional intensities into acoustic model training to improve the performance of expressive speech synthesis. (3)We proposed an automatic labeling technique of emphasis expression using a parameter generation technique of fundamental frequency to realize emphatic speech synthesis. (4)We proposed cross-lingual speech synthesis using only a target speaker's native language speech samples to synthesis multi-lingual speech at a low cost.