| The most care question by investigators on the domain of speech synthesis is how to improve the naturalness of the synthesis speech. Existing methods can' t receive perfect synthesis effect what people want thought they can improve synthesis at a certain extent. One of the reasons impacting the quality of synthesis speech is that the prosodic rules are not perfect. People learned prosodic rules through studying the large speech warehouse to obtain natural data to learn well. At this time, data mining become the preferred method to study speech synthesis.Our research is supported by Hunan national fund research project which is named as Research on the method of Speech Synthesis based on Data Mining (Project No. is 033JJY3097).It uses series of methods in Data Mining to study prosodic rules in speech synthesis.In this paper, we first optimize prosodic parameter in the method about picking up the rules from the prosodic parameter. We put forward the process of distilling rules above the optimized prosodic parameter. We point out a new algorithm-HLApriori algorithm to fit the speech analysis better, it can obtain more interested rules that investigators want.Secondly, we study variational patterns from two-word phrases. We obtain the mapping relationship between pitch and time length by training neural network, then count the pitch we need to synthesize by the trained network. We receive better learning effect and result fit the voted rules.Lastly, to learn the change of the syllable prosodic rules in sentences, we use the method of clustering analysis to obtain typical pitch model. On the base of it, we put the data we learned into deep describes, and use more than one method such as the clustering analysis, decision tree, neural network, rough sets in data mining to study prosodic rules, we receive considerable experiment results.Depending on all above studies and experiments, we can conclude that using data mining distill prosodic rules in speech synthesis is viable.
|
PDF Full Text Request