| Speech synthesis is one of the most important points of human-computer interactive. In this paper, the techniques of acoustic modeling and parameter tying strategy in Chinese speech synthesis are studied. A Chinese TTS prototype system is developed based on a Japanese speech synthesis system ( galatea-Talk ). In this system, we have made some improvements to meet Chinese characteristics. The work of this paper include:1．HTK/HTS platform is studied and analyzed. Based on HTK, an effective method is implemented for acoustic model training.2．XIF is used as synthesis' base unit set. In the Context Dependent acoustic modeling, unit's environment information is added to the model. The Decision Tree based state tying strategy in the Context Dependent acoustic modeling is studied. Decision tree design methods are analyzed, the design of question set are discussed. Using state tying decreases the size of the model. After state tying, the number of state is 2．84% of the state number before state tying. Prosodic information is trained by apply state tying to the duration HMMs and pitch HMMs.3．The synthesis part of Galatea-Talk (JPN) system was studied and analyzed. We keep its parameter generation module and synthesizer based on MLSA. To meet the Chinese characteristics, we redevelop the HMM reading module and HMM searching module. A text analysis module is implemented, in this part How-Net is used for processing the polyphone word. A Chinese speech synthesis system (CN-galatea-Talk) is designed and implemented by combining all the work above. In the SUS dictation test, the word mean correct rate is 89．5%. In the naturalness test, the mean opinion score is 3．15．The resulting run-time engine of this system is very small: about 2．2M bytes. Compared with a similar system based on diphone, the score increase significantly. Because of its small size, CN-galatea-Talk is applicable to mobile devices and embedded system. |
