The Research On Dai Prosody Prediction Module Of Speech Synthesis

As an important part of machine intelligence,speech synthesis technology has met a wide range of application needs,and the naturalness of synthetic speech is the key factor to affect the popularization and application.In order to improve the naturalness of Dai speech synthesis,the paper adopts the machine learning model such as decision tree and support vector machine(SVM)to explore the ways and methods to improve the effects of the prosody prediction.The main works of the paper includes:Firstly,the paper introduces the background of speech synthesis technology,explains the significance of Dai speech synthesis,and then respectively introduces various modules of speech synthesis,that is to say,including text analysis module,prosody processing module and speech synthesis module.In the second chapter,With the combination of the Dai language' characteristics,the paper gives the structure a detail analysis,dividing them into Prosody words,the secondary Prosody phrases and the primary Prosody phrase from three level and predicts the primary Prosody of phrases(L3).Based on the research and comparison of two kinds of prosodic structure prediction methods,this paper proposes a better performance based machine learning method based on statistical models.Secondly,according to the characteristics of the Prosody of Dai language,the decision tree model and the SVM model are used to predict the Dai's L3.The paper describes the C4.5 and CART algorithms in the decision tree model in details,and puts forward the prediction method of Prosody structure based on the decision tree.At the same time,the SVM algorithm is also described in details in the paper,and the process of predicting the Prosody structure is introduced.Thirdly,in order to enhance the users' experience,a prospective prediction platform is designed.The forecast effects of several different models are tested on the experiment platform respectively,through which we can get immediate data.Result shows:the decision tree and the SVM model used in this paper are helpful to forecast L3 annotations effectively.It can meet the basic requirements of improving the naturalness of the speech synthesis system.Finally,the paper makes a summary on the result of the experiment,points out several problems needed to be optimized in attribute set and algorithm for optimization.And last,some suggestions and prospects are put forward for the Prosody prediction module of speech synthesis.