Abstract: In order to improve the quality of speech synthesis based on HMM,this paper discussed the different structure and parameters on the effect of DNN training and demonstrated the validity of DNN discriminating S/U/V.The paper finished the speech synthesis of DNN on the HMM synthesis system were converted to the original speech spectrum parameter.Then,we studied on temporal decomposition(TD) algorithm to get the parameters of conversion program,and for DNN training set up the conversion model and event with no conversion function resynthesis of event vectors.The experiment proves that DNN conversion spectrum synthesis is closer to the original spectrum,and the subjective evaluation shows that this method can effectively improve the synthesized speech quality.

Key words: HTS,DNN,Deep leaning,Voice conversion,Temporal decomposition