【作者】 徐俊； 蔡莲红； 吴志勇；

【Author】 XU Jun1+, CAI Lian-Hong1, WU Zhi-Yong1,21(Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China)2(Department of SEEM, The Chinese University of Hong Kong, Hong Kong SAR, China)

【机构】 清华大学计算机系；

【摘要】 随着各国交流的不断深入,多语种以及混语种语音合成已经成为信息沟通和人机交互中越来越重要的部分。针对多语种和混语种语音合成的关键问题和现状,本文设计并实现了一个通用并且扩展性很好的多语种语音合成研究平台THMTTS,并基于其灵活的系统框架提出了一个混语种语音合成的基本流程,对中英日韩四种语言的混语种语音合成进行了语种检测和语音合成的研究。为进一步提高混语种语音合成技术水平提供了可能。更多还原

【Abstract】 Nowadays, multilingual and mixed-lingual speech synthesis has become more andmore important in information communication across different nations. Towards the keyproblem and current status in such researches, a new multilingual speech synthesis platform -THMTTS - is proposed in this paper. In the first part, the system architecture is presented.THMTTS comprises of 3 parts: basic data structure definition part, which provides a generaldata structure and information logging mechanism; module definition part, which givesresearchers power to design and implement new algorithms for speech synthesis; CrystalSonic, the graphic user interface (GUI), also the main entry point for speech synthesis,encapsulates the observations for data flow, debug information, module management, as wellas handling file I/O and controlling wave-out device. We designed a Multi-level data structurewithout restricting the contents, and the GUI part is able to call the pre-defined enumerationmethod to iterate all the data stored and expresses it with different appearances, depending onthe data type. Logs are also available to be listed in the GUI, as well as outputting to files orother streams. Another feature of this system is the smart module composition. Modulesshould implement the same interface and be realized in dynamic linking library (DLL). At thesystem initialization stage, all the modules stored in the specific place will be loaded, and then,users can manually choose which of them to be used and set the linking order. In the secondpart, multilingual and mixed-lingual support will be discussed. THMTTS aims to providespeech synthesis with language detection for 4 different languages including Chinese, English,Japanese and Korean. The modular structure itself has advantages for multiple languagesupport. The current system also integrated modules that carry out encoding conversion andlanguage detection. Language detection is based on Unicode, which is a general encoding forinternational use. The paper also proposed a statistical method based on the sum ofprobabilities to detect different language, which is proved to be effective by the experimentresult. In conclusion, the platform provides general and flexible system architecture forspeech analysis and synthesis. Based on this, a basic flowchart for mixed-lingual languagedetection and speech synthesis is introduced. The proposed architecture makes it possible toimprove the quality of mixed-lingual speech synthesis.更多还原