汉语普通话声调发音错误检测  

Mandarin tone error detection

作　　者：张琰彬[1] 呼月宁[2] 初敏[3] 黄超[3] 梁满贵[1] 

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]西北工业大学计算机学院,西安710072 [3]微软亚洲研究院,北京100080

出　　处：《清华大学学报（自然科学版）》2008年第S1期683-687,共5页Journal of Tsinghua University(Science and Technology)

摘　　要：汉语是声调语言,相同的音节带上不同的声调所代表的语义就很不相同。为了便于计算机辅助语言学习或用于普通话水平测试系统,准确地检测出声调的发音错误,该文采用精细的上下文相关的声调建模(contextdependent tone model,CDTM),并通过度量与实际发音最相符合的声调模型与预期的声调模型间的KL散度(Kullback-Leibler Divergence,KLD)来检测声调发音的正确性。实验结果表明,在控制错误接受率和错误拒绝率相等的前提下,错误接受率约为6.7%。Mandarin is a tonal language.Syllables having the same initials and finals,yet different tones may carry rather different meanings.When someone learns Mandarin,pronunciation of the tone is one of the most difficult parts.Therefore,computer assistant language learning systems need to accurately detect tone errors in speech.This paper describes a context dependent tone model to capture the variations of tones in different contexts with a Kullback-Leibler divergence(KLD) to relate an expected tone model for a ...

关 键 词：声调建模 发音错误检测 KL散度(Kullback-Leibler divergence) 

分 类 号：TN912.34[电子电信—通信与信息系统]