腭裂语音中齿龈塞音的声门代偿现象声学分析与判定  

Acoustic analysis and detection of glottal stops substituted for alveolar stops in cleft palate speech

作　　者：肖彦[1,2] 冯勇强[2] 赵庆卫[2] 马莲[3] 钱靖[3] 颜永红[2] 

机构地区：[1]北京交通大学信息科学研究所,北京100044 [2]中国科学院声学研究所语言声学与内容理解重点实验室,北京100190 [3]北京大学口腔医院,北京100081

出　　处：《声学学报》2015年第2期285-293,共9页Acta Acustica

基　　金：国家自然科学基金(60875014;60772039;10925419;90920302;61072124;11074275;11161140319;91120001;61271426);中国科学院战略性先导科技专项(面向感知中国的新一代信息技术研究;XDA06030100;XDA06030500);国家863计划(2012AA012503);中科院重点部署项目(KGZD-EW-103-2)资助

摘　　要：对齿龈塞音在腭裂语音中的声门塞音代偿现象进行了声学分析,计算频谱分布的多阶统计量—谱矩,并将代偿塞音和正常塞音进行对比。结果显示声门塞音爆破段的第一阶谱矩即频谱质心的频率位置比正常塞音低,因为声门塞音的阻塞部位在声门,导致声道腔体偏长从而共振频率偏低。还观察到声门塞音的第二阶谱矩即标准偏差偏高,说明其谱能量分布比正常塞音更加分散。声门塞音的第三阶谱矩即偏度大多为正值,反映了声门塞音功率谱的非对称性且大头朝向低频区而长尾朝向高频区。采用逻辑回归模型进行样本分类,通过交叉验证选出最优的四阶谱矩作为模型自变量,分类正确率为89.7%。结合塞音爆破时刻自动检测,实现了音节/di/的声门塞音客观判定。Glottal stops in cleft palate speech of Putonghua are abnormal productions in which closures in the glottis substitute typical obstructions along the vocal tract （e.g., alveolus, lips）. In this study, we investigated the spectral characteristics of glottal stops that patients intended to substitute for initial unaspirated alveolar stops. The first four moments were calculated for the burst and the subsequent early portion of the stop consonant. The results show that the first moment （i.e., centroid） was lower for glottal stops than for normal alveolar stops. That is, there was less acoustic energy in the high frequencies for glottal stops than for alveolar stops, due to the larger cavity anterior to the closure at the glottis than at the alveolus. We also found that the spectral energy of glottal stops spread over a wider frequency range, indicated by the greater second moment （i.e., variance or standard deviation）. Positive skewness showed that the spectral distributions were asymmetric with the head towards low frequencies and the tail towards high frequencies. As these spectral moments were taken as variable candidates, a logistic regression model was adopted as the binary classifier to determine whether an unaspirated alveolar stop consonant was a glottal stop. Acceptable classification accuracy was obtained after k-fold cross-validations and an optimal choice of a feasible to implement a system prototype to automatically detect classifier with the measurement of burst time. few spectral moments. This study suggests that it is glottal stops in cleft palate speech by integrating the

关 键 词：腭裂语音 声学分析 分类正确率 能量分布 逻辑回归 偏度 模型选择 功率谱 非对称 代偿性发音 

分 类 号：R767.92[医药卫生—耳鼻咽喉科]