Abstract: Automatic prosodic break detection and annotation arc important for both speech understanding and natural speech synthesis. We developed complementary model to detect Mandarin prosodic break by using acoustic, lexical and syntactic evidence. Our proposed method has the following advantages: (1) We do not adopt the independent assumption between the acoustic features and the lexical and syntactic features. (2) The complementary models not only in the features of the current syllable but also in the contextual features of the current syllable at model level realizes the complementarities by taking the advantages of each model. We verified our proposed method in a speech corpus of Chinese dis- course(ASCCD)，where 90.34% prosodic break detection precision rate can be achieved and 6.09% is improved than the baseline.

Key words: Prosodic break, Complementary model, Boosting classification and regression trec(CART) , Conditional random fields(CRFs) , Neural network(NN) , Support vector machine(SVM)