Chinese Prosodic Phrases Recognition Based On Syntax And Dependency

In the field of human-computer interaction,speech synthesis technology has always played an important role.How to improve the level of speech synthesis has become an important content in the present research.In the speech synthesis system,the level of speech recognition depends on the two elements of naturalness and intelligibility.Today,the first element of the intelligibility has reached the desired goal,while the naturalness of the other element is far away.The main performance is: the sentence of the machine output is poor,the language expression lacks priority,the machine smell is more serious.Therefore,improving the speech nature is the key point in the field of human-computer interaction.It can be seen from the module function of the text-to-speech conversion system that the accurate division of rhythm is the most important factor affecting the natural degree.Thus,this paper starts with the rhythm structure,and focuses on improving the accuracy of metrical structure division.High natural speech formation depends on the rhythmic hierarchy of high accuracy.For this,in terms of improving the natural tone of speech,this article embarks from the deep text characteristic,analyzes the relations between syntactic structure and rhythm structure,and the relationship between interdependence syntactic structure and rhythm structure,and then extract the text characteristic model training to improve the level of prosodic phrase of prediction.The main contents of this paper are as follows:(1)Feature analysis and extraction of syntactic structure and dependent syntactic structure.At present,the division of the rhythmical hierarchy is mostly based on shallow text features,such as the content of the word,the part of speech and the long of word,which is not used for the deeper syntactic features.However,deeper grammatical information or semantic information plays an important role in the accurate classification of prosodic phrases.Therefore,this paper uses the BerkeleyParser and LTP syntactic analyzer to analyze the syntactic data respectively,and then extracts the relevant feature items using the correlation algorithm.Finally,the best collection of each model is selected for the training and prediction.(2)Analysis of the relationship between dependency and syntax with prosodic structure.This article embarks from the syntax structure and interdependence syntactic structure,using the syntax analyzer respectively to extract the syntactic level,dependent type,connection,and the inner arc span of several characteristics to statistical analysis the relationship between the above characteristics and prosodic phrase boundary.Finally,these deep text features are combined with the above shallow text features to construct feature sets.(3)prosodic phrase recognition based on syntactic and dependent features.Based on the features of shallow text,this paper adds deep text features and builds different models based on these feature set.the feature set is used to construct the feature template,and then the CRFs,MaxEnt and BP-adaboost models are used to train and predict the different prosodic phrase prediction system.Finally,the results are compared with the prediction results of shallow text features.