基于维度模型的情感语音建模及生成方法研究

Speech is one of the most important approaches for human's communication. Not only linguistic information is conveyed in speech, but also speaker's emotional state. Up to date, the research on emotional speech modeling and generation mostly focuses on those typical and discrete emotion categories. From emotional speech modeling aspect, it's still far away from a full-scale emotional speech modeling. From application aspect, it also can not satisfy the need of generating human-like and flexible emotional speech for natural human-computer interaction purpose. Therefore, a research on dimensional model based emotional speech modeling and generation is selected as the topic of this research project. Building a full-scale emotional speech model and generating speech with a highly controllable emotional state is the target to achieve. In this research, the dimensional model of emotion is adopted to represent emotional state. On the basis of study on the labeling scheme of affective dimensions, and a comprehensive analysis on those emotion sensitive speech feature parameters and context features, a full dimensional space based emotional speech model is further established. An emotional speech generation method based on this model is also proposed. Finally, a prototype system for emotional speech generation based on the proposed model is built. A reasonable evaluation method for the generated emotional speech is also proposed. The work will greatly advance the development of research on harmonious human-computer interaction, speech understanding, and language cognition. It also possesses a wide application perspective.

语音是人类交流的最重要工具之一。人类的话语不仅起着表字达意的作用，而且还包含了说话人的情感状态等信息。目前情感语音建模和生成的研究基本集中在一些典型、离散的情感类别下进行，从情感语音建模层面而言没有达到完整的情感语音建模水平，从应用上而言也远不能满足自然人机交互中输出具有类人的、灵活多变的情感语音的需求。为此，本项目以基于维度模型的情感语音建模和生成为研究目标，尝试建立更为完整的情感语音模型，并能生成情感状态细微可控的语音。本项目采用维度模型进行情感状态的表征，在标注方法研究、以及对具有情感区分性的语音特征参数和上下文特征全面分析的基础上，建立完整情感维度空间上的情感语音模型，同时提出结合该模型的情感语音生成方法，最终建立一个基于维度模型的情感语音生成原型系统，及一个合理的情感语音评价方法。此项研究对推进和谐人机交互研究、语音理解和语言认知的发展将起到重要作用，同时还将有着广阔的应用前景。

语音是人类交流的最重要工具之一。人类的话语不仅起着表字达意的作用，而且还包含了说话人的情感状态等信息。目前情感语音建模和生成的研究基本集中在一些典型、离散的情感类别下进行，从情感语音建模层面而言没有达到完整的情感语音建模水平，从应用上而言也远不能满足自然人机交互中输出具有类人的、灵活多变的情感语音的需求。为此，本项目以基于维度模型的情感语音建模和生成为研究目标，尝试建立更为完整的情感语音模型，并能生成情感状态细微可控的语音。本项目采用维度模型进行情感状态的表征，在标注方法研究、以及对具有情感区分性的语音特征参数和上下文特征全面分析的基础上，建立完整情感维度空间上的情感语音模型，同时提出结合该模型的情感语音生成方法，最终建立一个基于维度模型的情感语音生成原型系统，及一个合理的情感语音评价方法。此项研究对推进和谐人机交互研究、语音理解和语言认知的发展将起到重要作用，同时还将有着广阔的应用前景。