Using large speech corpus (ASCCD) labeled with prosodic structure, the dissertation investigated the relationship between acoustic correlates and prosodic structure (involving prosodic boundary and stressed words) for Chinese Mandarin.We proposed automatic recognition model for prosodic boundary and stressed words. It explored the bridge from °at speech with combination syllables to structured speech with prosody information. We also studied the method for speaker adaptation by MAP algorithm. Finally, we analyzed the correlativity between prosodic structure and syntax structure.The dissertation includes the following work:(1) We studied the acoustic correlates for prosodic boundary and stressed words. By statistical method with large speech corpus, it validated some important argumentations in phonetics such as declination of pitch contour in a phrase,pitch resetting on boundary, prosodic structure correlated with bottom line and stressed words correlated with top pitch line separately, and so on.(2) In the dissertation, we constructed a series of feature vectors, and proposed a CART model for prosodic boundary location. Such model could also evaluate the importance of each feature vector.(3) Thinking about the prosody hierarchy, we proposed multi-step model for prosody boundary location, the GMM model also be tried. Both improved the performance.(4) Further, we used CART model for stressed words detecting, and a mixture model with probability was also proposed.(5) We investigated speaker adaptation method with MAP algorithm, which improve the models' robustness both for prosody boundary location and accent words detection.(6) In the end, we analyzed the correlativity between prosodic structure and syntax structure. And managed to combine the syntax information to improve the accuracy of boundary location.
修改评论