基于音素评分模型的发音标准度评测研究  

Pronunciation Quality Evaluation Based on Phone Scoring Model

作　　者：严可[1] 戴礼荣[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：《中文信息学报》2011年第5期101-108,共8页Journal of Chinese Information Processing

基　　金：国家十五重点项目资助(ZDI105-B02)

摘　　要：在计算机辅助语言学习系统中,后验概率是普通话水平测试（PSC）电子化系统衡量考生发音标准程度的重要指标,但后验概率与人工的主观评分存在着显著差别。该文提出了＂音素评分模型＂的思想,对后验概率进行变换。该文研究了线性和非线性的sigmoid音素评分模型,并发现线性音素评分模型有闭式全局最优解,非线性音素评分模型可用梯度下降法求解。在全国采集的498人的普通话考试现场数据集上的实验表明该策略能使系统评分性能有明显的提升：当后验概率在全音素概率空间中计算时,可使系统性能提升约42%;当后验概率在优化的概率空间中计算时,能使系统性能提升约23%~27%。Posterior probability is a promising feature for computers to judge testers＇ pronunciation quality in computer assisted language learning systems.However,the discrepancy between posterior probability and evaluators＇ criteria is obvious.This paper introduces ＂Phone Scoring Model＂ which transforms posterior probability to deal with the problem.Both linear and non-linear phone scoring models are investigated and we find that： close formed solution can be obtained for linear phone scoring models and gradient descent method can be used for nonlinear phone scoring models.Experimental results based on 498 people＇s live PSC database indicate that this approach can significantly improve system performance： approximately 42% relative performance gain when posterior probabilities are calculated with all-phone probability space;approximately 23%~27% relative performance gain when probabilities are calculated with optimized probability spaces.

关 键 词：语音评测 音素评分模型 后验概率 普通话水平测试 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]