辅助语音评分系统中一种流利度自动评分方法  

Automatic assessment of speech fluency in computer aided speech grading systems

作　　者：黄申[1] 李宏言[1] 王士进[1] 梁家恩[1] 徐波[1,2] 

机构地区：[1]中国科学院自动化研究所数字内容技术研究中心,北京100080 [2]中国科学院自动化研究所,模式识别国家重点实验室,北京100080

出　　处：《清华大学学报（自然科学版）》2009年第S1期1349-1355,共7页Journal of Tsinghua University(Science and Technology)

基　　金：国家"八六三"高技术项目(2006AA010103)

摘　　要：流利度评分是目前计算机辅助自动发音评分系统中一个重要组成部分。为了解决流利度特征表述中和内容、韵律等高级表述技巧相关的广义流利问题,该文提出了一种能够全面客观地评价口语流利度通顺性和韵律性的评分方法。该方法除了通顺性之外,可以提取停顿、韵律、连读和失去爆破等特征;评分过程中,系统采用多层次融合方法,分别从句子级和篇章级提取流利度特征,其中篇章级得分由句子级加权得分和篇章级特征进行融合得到。在实验中,分别以机器得分和人工平均分的相关度、均方差等指标比较了多元线性回归(LR)、BP神经网络、支持向量回归机(SVR)3种算法的性能。结果表明:基于非线性建模的BP神经网络和支持向量回归机拟合的得分模型要好于多元线性回归;该方法可以作为一个重要指标,应用在计算机辅助语音评分系统中。A speech fluency grading system for general representation of fluency was developed to evaluate both smooth reading and more advanced skills related to content-based coherent expression and rhythmic proficiency.In addition to traditional prosodic features,this approach also analyzed pause,rhythm,links and assimilation skills in fluent English pronunciation.The scoring stage uses a hierarchical scoring fusion strategy with sentence and passage level using fusion measures such as multiple linear regression(LR),back propagation neural networks(BP),and support vector regression(SVR).The correlation and mean square errors between the ground-truth estimated scores and automatically estimated scores suggests that the non linear score fitting methods(BP,SVR) are superior to the linear method.Thus,this approach should be adopted in computer aided automatic scoring systems to improve efficiency and accuracy.

关 键 词：计算机辅助语言学习 语音识别 流利度 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]