Research Of Text Analysis And Prosody In Chinese TTS System

Text-to-Speech (TTS) technology aims to transform the text information into speech signal with the rules of speech processing and make the computer read the text information to let us understand the meanings from listening. In this thesis, auto word segmentation, prosodic labeling and prosodic construction prediction are presented.In general, TTS system consists of text analysis, prosody control and speech synthesis. Text analysis module is important in TTS system. It simulates the comprehended process to natural language of people, make the computer understand the input text contents and give the pronunciation hints to the latter modules. In it, segmentation, phonetic notation and part-of-speech tagging are primary important components of text analysis, and they are also issues which need to be resolved .With the ambiguous words and unknown words, Chinese auto word segmentation becomes the main problem of Text-to-Speech system. N-gram is a word segmentation algorithm based on statistics. Compared with the other algorithm, it has a better performance in ambiguous words segmentation, but it is not enough. The thesis presents an improved-mixed strategy which based on N-gram combined with the maximum matching pretreatment, the new algorithm also gets united with the POS disambiguation and smoothing part. The experiment shows that the correct ratio and recall ratio are improved.Prosodic processing generalizes the super-segmental features including pitch, duration and energy, it makes the output can express exactly and naturally. The result of text analysis is syntax words which is not equal to prosodic words. So prosodic processing is needed. Prosodic features such as tone, rhythm and stress can be showed by the variation of super-segmental features. Therefore, the changes of these features become the base of prosodic controlling. On the basis of XML, a Chinese prosodic labeling language is presented to tag the output of prosodic analysis automatically and make the output of TTS system naturally.In order to express ourselves and understand others, we must know well all kinds of boundary features of prosodic units, divide the prosodic hierarchical boundary and choose the prosodic units correctly, as to know the prosodic construction. At last we discuss the acoustics behavior of the boundaries by experiments. Combined with the key auxiliary word, the CART-tree model improves the precision of the prediction and makes the text analysis module links with the prosodic processing module.