基于PLDA的说话人识别时变鲁棒性问题研究  

Research on time-varying robustness in speaker recognition based on PLDA

作　　者：陈霄鹏[1] 彭亚雄[1] 贺松[1] 

机构地区：[1]贵州大学大数据与信息工程学院,贵州贵阳550025

出　　处：《微型机与应用》2016年第5期58-60,64,共4页Microcomputer & Its Applications

摘　　要：随着时间的变化,人的声音也会发生变化。这对说话人的识别带来了一定的影响。通过研究发现,说话人识别的性能与时间有着线性变化的规律。传统的说话人识别系统使用GMM-UBM模型并不能很好地学习出线性变化规律。由于概率线性判别分析(PLDA)对于类内与类间有着很好的线性区分度,所以为了解决线性变化的问题,选择概率线性判别分析的方法学习说话人识别中时变的线性变化规律。从实验结果看出,PLDA对于说话人识别的识别鲁棒性具有很好的提升。As time goes on,the voice will have a change. It is an influence to speaker recognition. By our research,we find that the recognition rate of speaker recognition have some rule of linear on time-varying. The traditional speaker recognition system always uses GMM-UBM,but it can＇t learn the rule of linear. The Probabilistic Linear Discriminant Analysis（ PLDA） can distinguish intra-class and inter-class easily.So in order to solve the linear problem,we choose PLDA to learn the rule of speaker recognition on time-varying. The experiment results show that PLDA is better for time-varying robust in speaker recognition.

关 键 词：说话人识别 时变鲁棒性 GMM-UBM PLDA 

分 类 号：TN912.34[电子电信—通信与信息系统]