语调短语预测中长度约束模型的对比研究  

Comparative study of the length constraint models for intonational phrase prediction

作　　者：周游[1] 刘方舟[2] 

机构地区：[1]湖南财政经济学院应用数学系,长沙410205 [2]湖南师范大学数学与计算机科学学院,长沙410081

出　　处：《清华大学学报（自然科学版）》2013年第6期787-790,795,共5页Journal of Tsinghua University(Science and Technology)

基　　金：湖南省自然科学基金项目(13JJ6030);湖南财政经济学院科学研究项目(K201010;K201112)

摘　　要：从文本中准确地预测语调短语是提高合成语音自然度的关键之一。已有的语调短语预测方法多从词法或浅层句法信息入手,然而自然语流中语调短语的划分不仅与语法结构相关,也受其长度分布规律的约束。该文详细地比较了基于规则和基于概率、嵌入式和后接式、全局最优和局部最优等多种语调短语长度约束模型。实验结果表明:口语中语调短语的规划是一种短时的局部规划,并且对短语长度独立建模的效果优于将短语长度信息直接加入统计分类模型的特征集。Accurately predictions of the intonational phrases from a text are important for improving the naturalness of synthetic speech.Current intonational phrase predictions mainly use lexical or shallow syntactic information.However,the organization of intonational phrases in normal speech is not only related to the grammatical structure,but is also influenced by the phrase length distribution.This paper compares a variety of length constraint models for intonational phrases.The results show that intonational phrase planning involves short-term local planning,and that independent modeling of the phrase length performs better than directly adding phrase length into the feature set of a statistical classification model.

关 键 词：韵律结构 语调短语 长度约束 最大熵 

分 类 号：TN912.33[电子电信—通信与信息系统]