一种利用主曲线的说话人自适应方法  

Speaker adaptation method using principal curves algorithm

作　　者：王晶莹[1] 王作英[1] 

机构地区：[1]清华大学电子工程系语音识别实验室,北京100084

出　　处：《高技术通讯》2007年第5期470-473,共4页Chinese High Technology Letters

基　　金：863计划（2001AA114071）资助项目.

摘　　要：为了克服噪声对语音识别系统的影响,提出了一种基于主曲线的说话人自适应方法,这种方法可以通过一组主曲线描述所有状态的特征统计信息和码本参数之间的关系,并利用特征统计量在主曲线上的投影更新码本。当有背景噪声存在时,这种方法可以有效修正状态的特征统计信息以减弱或去掉噪声的影响。在863大词汇量连续语音识别数据库上的实验结果证明：这种方法相对于基线以及本征音说话人自适应算法,系统识别性能都有明显的提高。This paper proposed a new speaker adaptation method utilizing principal curves algorithm to reduce noise＇ s effect on speech recognition systems. The key feature of this method was the construction of principal curves describing the correlation between observations of different acoustic states and codebook mean. Herein the projection of feature statistics on principal curves was taken as the updated codebook parameters. When noise exists, the method can modify effectively feature statistics and reduce noise＇ s effect. The results of the experiment on the 863 large vocabulary continuous recognition database showed that this new method is superior to Baseline and EigenVoice

关 键 词：主曲线 说话人自适应 相关性 空间相关性 

分 类 号：TN912.34[电子电信—通信与信息系统]