基于HMM语音合成的语调控制  

Control of intonation in HMM based text-to-speech systems

作　　者：王永鑫[1] 贾珈[1] 张雨辰[1] 蔡莲红[1] 

机构地区：[1]清华大学计算机科学与技术系,清华信息科学与技术国家实验室,普适计算教育部重点实验室,北京100084

出　　处：《清华大学学报（自然科学版）》2013年第6期781-786,共6页Journal of Tsinghua University(Science and Technology)

基　　金：国家"九七三"重点基础计划项目(2013CB329304);国家"八六三"高技术项目(2012AA011602);国家自然科学基金项目(61003094)

摘　　要：语调是语音分析和合成领域关注的重要课题,可计算的语调模型是实现语调控制的关键。该文分析了大规模语句的音节音高在句子中的变化,归纳了语调模式。陈述语调主要表现为基调的升降和音高下倾;疑问语调主要特点是疑问焦点的音高提升和调型变化。该文提出了一种陈述语调归一化描述方法,以及疑问语调调型差异模型。利用基于隐Markov模型的语音合成系统的控制机制,实现了对语调的控制。实验表明:基于陈述句语调归一化描述方法模拟了陈述句语调的变化,基于疑问语调调型差异模型实现了陈述到疑问语调的转换。语调控制使合成语音的表现力得到了提高。Intonation is an important research topic in speech analysis and synthesis,with a computable intonation model as the key to achieving intonation control in synthesized speech.This study analyzes the pitch movement of syllables in a sentence to describe the intonation patterns in Chinese sentences.Declarative sentences have downward pitches while interrogative sentences raise the pitches and have more elaborate pitch contours.A unified intonation representation is then given with an intonation pattern difference model for interrogative and declarative sentences.The model is used to control sentence intonation in a hidden Markov model（HMM） for speech synthesis.Tests show that this unified intonation representation can produce the different declarative intonations and the intonation pattern difference between interrogative and declarative intonation.More expressive speech is then generated with the intonation control.

关 键 词：语调控制 语调 基于隐Markov模型(HMM)语音合成 语调模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]