Chinese Natural Spoken Rhythm Analysis And Automatic Label Study

With the rapid development of computer technology in recent years, great progresses have been made in phonetics and speech technology. More and more speech products facilitate people's life, such as speech input system, human-machine dialog system, computer-aided language studying system, and so on. However, speech science is still confronted with many challenges. A lot of new problems have arisen in the research from isolated words to continuous speech, especially prosody.Past research on prosody is mostly based on read speech and small corpora. But recent research shows that there are lots of differences between read speech and spontaneous speech. It is spontaneous speech that can reflect the prosodic characteristics of natural speech. Therefore prosodic research should be based on a large spontaneous speech corpus.In this thesis, at first, the author selected about 4 hours of the recorded conversations from CADCC (a large spontaneous speech corpus recorded in ordinary rooms) automatically. Secondly compared with other read speech corpora, the differences between spontaneous speech and read speech are discussed. Thirdly, based on the selected corpus, all kinds of prosodic characteristics are analyzed by using statistical methods to get the main cues of prosodic boundaries and stresses in spontaneous speech are achieved. Finally, the decision tree method is applied to build an automatic prosodic-labeling system on the basis of the statistical results to help speed up prosodic annotating.Furthermore, three new ideas are involved in the thesis: first, spontaneous speech is regarded as the research material which is rarely considered in past studies; second, a rule-based statistical method is applied on phonetic research, which is a new research thought; third, the research is based on a large corpus and the results are concluded by a great deal of statistical analyses, which is a latest method in research on linguistics.