Research On Prosodic Control In Mandarin-Tibentan Cross-language Conversion

In recent years, with the development of the human machine speech interface, voice conversion has been the research focus, and being gradually applied to education,communication, and many other fields. In our country, as the representative of the mainstream,the method of mandarin has made great progress, and been becoming mature. However,nowadays, there are little researches on cross-language voice conversion of minority languages, especially the study of Mandarin-Tibetan cross-language conversion is more limit.At the present stage, building preliminary a Mandarin-Tibetan cross-language voice conversion system, this system only have a simple transform, and no prosodic control. What's more, on the synthetic stage, according to the needs of practical application, the initials and the finals of Tibetan Lhasa dialect are selected as speech synthesis unit, and only having a simple concatenative synthesis with it, so the nature degree of coverted speech is not well. So based on the above situation, this paper researches on the method of prosodic control of Mandarin-Tibetan conversion. meanwhile, selecting the Tibetan syllables as unit. The experimental results show that the study achieves the purpose of improving the speech naturalness and the quality of speech. The thesis' s main work and innovation are as follows:Firstly, building a Tibetan Lhasa dialects corpus that is used to implement the Mandarin-Tibetan cross-language voice conversion, and a library of Mandarin prosodic feature analysis for prosodic control. The main feature of the Tibetan syllable library is containing the different syllables of different location. Chinese metrical library contains different tonetic individual words,and different types of prosodic words and phrases.Secondly, studying the prosodic features of the mandarin and building a prosodic model.According to the method named five-degree tone value andstatistical methods to build a fundamental frequency model for mandarin as well as a transformation model of syllable duration and pause duration.Thirdly, this paper implements the Mandarin-Tibetan cross-language speech conversion based on prosodic control, the syllables of Tibetan Lhasa dialect are selected as speech synthesis unit to synthesize mandarin speech with waveform concatenation. Under the guidance of founded metrical model, modifying the acoustic parameters to the obtained speech of mandarin through voice conversion, including the fundamental frequency,duration,and so on, then making prosodic control.Finally, evaluating the speech that has been transformed through Mandarin-Tibetan cross-language speech conversion based on prosodic control. Experimental results show that the sound quality and the naturalness of Mandarin-Tibetan conversion after prosodic control is effectively improved, achieving the purpose of the study.