基于发声模型的腭裂语音高鼻音自动检测算法  

Automatic detection algorithm of hypernasality in cleft palate speech based on acoustic model

作　　者：张桠童 何凌[1] 张婷[2] 尹恒[3] 李杨[3] 

机构地区：[1]四川大学电气信息学院,四川成都610065 [2]成都航空职业技术学院国有资产管理处,四川成都610100 [3]四川大学华西口腔医院,四川成都610041

出　　处：《计算机工程与设计》2015年第6期1592-1597,共6页Computer Engineering and Design

基　　金：国家自然科学基金青年基金项目(30900391)

摘　　要：通过对腭裂语音发声模型进行研究,提出基于激励、声道、辐射模型特征参数的腭裂语音高鼻音等级自动识别算法。通过对基于激励模型的基音频率、基于声道模型的共振峰参数、基于综合发声模型的短时能量和Mel倒频谱系数等表征高鼻音特性的参数进行分析和改进,以K-最近邻算法为模式识别分类器,得出应用4种特征参数的高鼻音等级自动识别结果。实验结果表明,Mel倒频谱系数与腭裂语音临床生理特征相关性最大,对不同等级高鼻音识别率最高。Though the studies on the acoustic model of cleft palate speech signals ,the automatic identification algorithm of hy‐pernasal level in cleft palate speech based on characteristic parameters of incentive model ,vocal tract model and radiation model was put forward .After a series of analysis and improvement on characteristic parameters of hypernasality including pitch based on incentive model ,formants based on vocal tract and energy and Mel cepstrum coefficient based on comprehensive acoustic pa‐rameters ,four kinds of recognition results on characteristic parameters of hypernasality levels were obtained by using K‐nearest neighbor .The experimental results show influenced by cleft palate ,the best recognition rates is gained by using identification al‐gorithm of Mel cepstrum coefficient ,which has maximum correlation with cleft palate speech in clinical characteristics physically .

关 键 词：腭裂语音 数学模型 基音频率 共振峰 能量 Mel倒频谱系数 K-最近邻算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]