Research On The Method Of Mandarin-tibetan Bilingual Conversion

In recent years, with the rapid development of the human voice interaction technology, voice conversion technology has been attention by many researchers, and will be gradually applied to many fields such as education and communication. In the domestic, to the mainstream languages such as mandarin and cantonese, the research of speech conversion method has made great progress. But it is lack of the cross-language speech conversion system of minority dialects. Tibetan is one of the ancient minorities in our country, and the large number of people speak the Tibetan, wide distribution area.In this thesis, the Tibetan Lhasa dialects as the research object, it establishes 2800 Tibetan Lhasa dialects corpus, segmentation and annotation acoustic finals information and the establishment of a Tibetan acoustic finals library.In Mandarin-Tibetan bilingual cross-language speech conversion, the first need to translate the text of the needing translation Tibetan into corresponding Chinese text, put the Chinese text into text analysis to obtain all the acoustic finals, and then find the acoustic finals library who good indexes has been established, In Tibetan acoustic finals for primitives, using the boundary information at the same time, according to the context of problem sets and candidate primitive spectrum distance for the establishment of a Classification and Regression Tree. For the goal of Chinese sentences, it uses the Classification and Regression Tree algorithm to choose the most conforms to the context information of the acoustic finals, selecting the location of pronounce and the best quality of the acoustic finals, and then respectively uses waveform concatenation synthesis and STRAIGHT algorithm to get the corresponding Chinese phonetic statements, so as to completing research on the method of Mandarin-Tibetan bilingual conversion. The thesis’ s main work and innovation are as follows:Firstly, this thesis establishes 2800 Tibetan Lhasa dialects corpus, extracts and establishes a Tibetan acoustic finals library. Firstly, it designs Tibetan language text corpus, and then to recorded speech corpus for segmentation and tagging for all acoustic finals information, finally according to the Tibetan language acoustic finals classifies and create indexes. Thus the Tibetan voice finals library is established in this thesis, it lays a solid foundation for the Sino-Tibetan bilingual cross-language speech conversion.Secondly, the STRAIGHT algorithm is adopted in the Mandarin-Tibetan bilingual cross-language speech conversion. It can be very flexible to modify the fundamental frequency of speech signal, cycle index and smooth spectrum related parameters, so as to improve the quality of target voice.Thirdly, this thesis implements the Mandarin-Tibetan bilingual cross-language speech conversion. For the goal of Chinese sentences, it uses the Classification and Regression Tree algorithm to choose the most conforms to the context information of the acoustic finals, selecting the location of pronounce and the best quality of the acoustic finals, and then respectively uses waveform concatenation synthesis and STRAIGHT algorithm to get the corresponding Chinese phonetic statements, and it performs MOS evaluation, DMOS evaluation and ABX test. Experimental results show that the sound quality of the conversion of using STRAIGHT algorithm is better than using waveform concatenation synthesis.