面向情感转换的层次化韵律分析与建模  

Hierarchical prosody analysis and modeling for emotional conversions

作　　者：徐俊[1] 蔡莲红[1] 

机构地区：[1]清华大学计算机科学与技术系普适计算教育部重点实验室(清华大学),北京100084

出　　处：《清华大学学报（自然科学版）》2009年第S1期1274-1277,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(60805008;90820304);国家"八六三"高技术项目(2007AA01Z198);国家"九七三"重点基础研究项目(2006CB303101)

摘　　要：语音的韵律特征对情感表达起着非常重要的作用。在基于韵律修改的情感语音转换系统中,为了更为准确地预测情感语音的韵律特征,该文提出一种层次化的韵律分析与建模方法,针对愤怒、高兴、悲伤、惊奇这4种情感对语音按照韵律结构的层级进行分析,探讨不同层次之间情感韵律特征变化规律的关联性和叠加性,并使用高层音段信息作为输入对低层信息进行决策树建模。实验结果表明,本方法对情感韵律特征的预测均方误差比局部分音段以及传统的决策树韵律模型低大约5%。Prosody features play an important role in emotional expressions in human speech.A hierarchical analysis and modeling framework was developed to predict the prosody of emotional speech for an emotion speech conversion system based on prosody modification.The prosody features were analyzed for angry,happy,sad and surprise according to the Mandarin prosody structure.Global features were taken into account while modeling the local features by a decision tree.Tests show that the result achieved about 5% lower root mean square error than global,local segmentation and traditional decision tree models.

关 键 词：情感韵律特征 情感转换 韵律分析 韵律建模 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]