Based Hmm Can Be Training Vietnamese Speech Synthesis System

Speech synthesis is the process that reappear the understandable voice signal through the special hardware equipment or the computer. The speech synthesis technology is one of key technologies in realizing the man-machine speech communication, and establishing a spoken language system which has both heard and oral ability. The research of Speech synthesis technology had more than 200 years history, but the latter-day speech synthesis technology which has the practical significance is truly develops along with the development of computer technology and the digital signal processing technology.Over the past few decades, methods of speech synthesis mainly have:Articulatory Synthesis, Source-filter Model Synthesis, Unit-Selection Synthesis and Trainable TTS. These methods have merits and drawbacks respectively, but comparatively speaking, Trainable TTS has higher degree of automation, and small dependence with the different pronunciation person, the different pronunciation style, and the different language. Based on these characteristics, this thesis has selected a trainable speech synthesis method based on the Hidden Markov Model (HMM) to carry on the construction of synthesis system.Vietnam located at the east of Indo-China Peninsula, Southeast Asia. It is bounded on Yunnan of China, which has brought the frequent exchange of language culture and related talented people of two places, as well as the region superiority of researching Vietnamese new phonetic technology. Therefore this thesis has studied the Vietnamese speech synthesis system, and hoped that the result of research can be used to practice and the Vietnamese man-machine interaction can be realized. The main jobs of this thesis are as follows:(1) Expounding the fundamental principle of HMM, and introducing the procedure of constructing a trainable speech synthesis system which based on HMM.(2) Introducing the phonetic features of Vietnamese, and reviewing the present situation of Vietnamese speech synthesis. On that basis, proceeding data preparation of vietnamese speech synthesis system. The data preparation work mainly includes: Constrution of Corpus, determination of phoneme tabulation, labelling training data as well as design of context attribute and question collection. The most important part of work is labelling the training data. In this thesis, we select a method that combines the handwork with the procedure to carry on corpus labelling.(3) In the Cygwin platform, according to the training process of STRAIGHT synthesizer, using 500 labeled sentences to accomplish the model training process and back-end synthesis part of vietnamese speech synthesis system. Generating waveform using synthesizer after rhythm labeling of synthetic sentences.The experimental result shows that the STRAIGHT synthesizer for Vietnamese speech synthesis is feasible. Subsequent work should be realized on the automatic analysis of Vietnamese texts, as well as improving the naturalness of synthetic speech.