摘要
在针对电话语音的自动语种识别系统中,训练和测试语料之间存在不同说话人、信道等因素差异带来的不匹配,是影响识别性能提高的关键因素。为了消除此类影响,提出一种层次化空间分析方法,首先对前端部分MFCC+SDC特征进行HLDA(异方差线性判别分析),增大了语种各个类的类间差异;然后对经自适应得到含有冗余信息的GSV进行PCA特征选择,有效地去除了信道等冗余信息的干扰。实验结果表明,此方法能有效消除信道等噪声影响,从而提升了原有系统的识别性能。
In automatic spoken language recognition system on telephone conversation speech,differences between train and test utterances on channels,gender and speakers are the key factor of improving the performance of the system.This paper proposed a hierarchical space analysis method.Firstly,it mapped the front-end cepstral features of SDC into the HDLA space,aiming at increasing the discriminability between different languages.Secondly,it selected the characters of adaptive GMM super vector by the method of PCA,which eliminated the influences of different channels,speakers and so on.Experiment results indicate that this method is better for improving the system＇s performance than the original baseline system.
出处
《计算机应用研究》
CSCD
北大核心
2012年第10期3651-3654,共4页
Application Research of Computers
基金
国家"863"计划重点资助项目(2008AA011002)
关键词
语种识别
层次化
空间分析
冗余信息
language recognition
hierarchical
space analysis
redundant information
