基于整句相似性计算的韵律短语预测模型  

New Prosody Phrase Prediction Model Based on Whole Sentence Similarity Computing

作　　者：李剑锋[1] 胡国平[1] 王仁华[1] 

机构地区：[1]中国科学技术大学讯飞语音实验室,安徽合肥230027

出　　处：《小型微型计算机系统》2006年第10期1935-1938,共4页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(69975018)资助.

摘　　要：语音合成系统中,韵律短语的预测对合成语音的自然度有重要影响.为了突破主流的基于决策树预测方法的若干缺陷,提出了基于整句相似性计算的韵律短语预测模型.通过对1000个句子的测试,该方法在可接受的语料手工标注工作量的范围内,超过了传统决策树的方法.In speech synthesis systems, prosody phrase prediction has great impact on the naturalness of synthesized speech. To avoid some disadvantages of prevailing decision tree based method, a new model for prosody phrase boundary prediction based on whole sentence similarity computing is proposed in this paper. Through tests on 1000 sentences, it can be proved that the new model has the ability of outperforming the decision tree based method with an affordable annotated corpus.

关 键 词：韵律短语预测 句子相似性 编辑距离 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]