An Hmm-based Speech Synthesis System Applied To Tianjin Dialect

Due to the extensive use in artificial intelligence, speech signal processing and human-computer interaction, speech synthesis has become the research focus for many companies and labs. With the gradual improvement of the produced voice, users make more and more demands for the system especially for the diversity of the synthesized speech. And dialect speech synthesis can not only be applied in more and more multimedia, but also increase the diversity of the system, which could be popular for many locals. Therefore, this paper put much attention on dialect synthesis principle, process and results based on the traditional HMM-based synthesis system. Specific studies of this paper are as follows.First of all, this paper makes a simple review of different methods and the history of speech synthesis. The system has undergone source-filter filter method, unit-selection method, HMM-based method and so on. After introducing individual's advantages and disadvantages, this paper also infers the importance of dialect speech synthesis to provide a starting point for the following paper.Then the article focuses on introducing the Hidden Markov Model and its use in speech modeling. After that, the features of Tianjin Dialect are analyzed in detail with combination with HTS system. Then modifications of the module in the system and the following research priorities are discussed.As the main difference between Tianjin Dialect and Mandarin is the tone change, two methods are designed in the following paper. For the first methods based on rules, three detailed methods are put forward——using Tianjin dictionary directly in the system, treating Tianjin Dialect as irregular Mandarin and using mapping rules between these two languages. For the third method, the paper also designs two methods based on different tone models. The optimal program based on rules is given through experiment.In the end, after analyzing the limitation of using rules for Tianjin Dialect, the paper also increases the tone precision by improving the HMMs through HTS system itself. In order to achieve this purpose, many ways are tried through experiments like using bilateral question set, extracting more context features and using moving tone window of question set. The effects of different corpora, dialects on the produced voice are also discussed. At last, the paper also gives prospect for Sichuan Dialect and Shanghai Dialect so that it could provide a basis for future work.