Research On Text-to-Speech Systems Based On SC-Analysis

Text-to-Speech(TTS) technique has close relations with Natural Language Processing and Speech Signal Processing. It tries to simulate human's speech ability by using the computer devices. The research on TTS will engender great influences on Human-Machine Communication, Natural Language Interface and Intelligent Computer Architecture, etc.In this thesis, the new techniques and algorithms about TTS were summarized and analyzed, and a new two phases model of TTS was presented. Based on this model, the now available methods of TTS will be unified with one pattern and hence it will accelerate the theoretic research on the fundamental problems of TTS.A new method for enhance the performances of TTS was put forward by using morphological analysis, parsing and context analysis. In this way, the segmental features and super-segmental features of a sentence were acquired, the normalization of input text, parsing and letter-to-pronunciation were implemented based on SC-analysis.In the process of Normalization of input text, we studied the problems such as the determination of sentence's border, the pronunciation of some special symbols, the expansion of acronyms or abbreviated words, and the determination of phrase's border. A new algorithm was proposed for the determination of sentence's border, which was based on structure features and context conditions. In addition, the special processing was carried out for the ambiguous punctuations, such as ",","." and ":". The pronounciation of digits was discussed. For the expansion of acronyms or abbreviated words, dictionary and context analysis scheme was put forward and realized. We also considered the determination of English phrase's border and Chinese word segmentation.We carried out the determination of English phrase's border by using phrase dictionary and phrase rules. In the research on Chinese word segmentation, the MM approach was improved to accomplish the segmentation and detection of ambiguous character strings by using MM approach only one time, and the method based on word frequency and length was introduced to implemented...