基于隐马尔可夫模型的非母语说话人口音识别  

Accent classification of non-native speakers based on hidden markov models

作　　者：赵征鹏[1] 杨鉴[1] 普园媛[1] 尉洪[1] 

机构地区：[1]云南大学信息学院,云南昆明650091

出　　处：《云南大学学报（自然科学版）》2004年第B07期43-46,共4页Journal of Yunnan University(Natural Sciences Edition)

基　　金：国家自然科学基金资助项目(60265001).

摘　　要：基于云南境内说话人母语为纳西语、傈僳语的汉语普通话语音,采用隐马尔可夫模型(HMM),由标准普通话语音库训练得到基线系统的声学模型(HMMSTD).然后以基线系统的声学模型为初始模型,分别用母语为纳西语、傈僳语的汉语普通话语音训练得到各自的声学模型(HMMNX和HMMLS).对于未知语音,根据各识别系统的概率得分,采用最大概率准则进行口音的分类判决.实验表明,基于HMM的口音识别系统,其正确识别率达93%.It is concerned Mandarin speech from speakers which natives are Naxi and Lisu in Yunnan province,using hidden markov models,and getting the phonetic models of the baseline system (HMM_(std)).Then referred to the phonetic models of the baseline system as the initialized models,with the non-native Mandarin speech from Naxi,Lisu,and trained their own accent models (HMM_(NX)&HMM_(LS)).For unknown accented speech,by means of the logarithmic scores associated with each system and the maximum likelihood criterion,automatic accent classification of non-native speech can be made.The experiment also showed that accent classification based on HMM could achieve the correct rate of 93%.

关 键 词：隐马尔可夫模型 非母语说话人 口音识别 基线系统 声学模型 民族语 

分 类 号：TN912.34[电子电信—通信与信息系统] O211.62[电子电信—信息与通信工程]