说话人识别中语速鲁棒性研究  

Research on speaking rate robustness in speaker recognition

作　　者：朱紫阳[1] 贺松[1] 彭亚雄[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7期54-56,共3页Microcomputer & Its Applications

摘　　要：如今,说话人识别技术已经比较成熟,但依然有很多因素影响说话人识别系统的稳定性。本文针对说话速度对说话人识别的影响进行了一系列的研究工作。通过模型空间分布可视化和语音频谱观察两方面来分析不同语速语音的差距。然后,提出了最大似然线性回归(MLLR)和Constraint MLLR(CMLLR)的方法对模型和特征进行变换,使训练端和测试端的语音特征互相接近匹配。通过实验发现,MLLR和CMLLR能较好地提高说话人识别系统中语速鲁棒性。Recently, speaker recognition has been matured, but there are still so many factors impact the sability of speaker recognition system. This paper mainly researches the influence of speaking rate on speaker recognition. Through making distribution of model space visualization and observing the print of frequency spectrum to analyse gap of the different speed voice. Then, we propose the method of Maximum Likelihood Leaner Regression （MLLR） and Constraint Maximum Likelihood Leaner Regression （CMLLR） to transform the model and feature. It is aimed at making training and testing mutual match. Through the experiment, we find that the MLLR and CMLLR can improve the robustness in speaker recognition with different speaking rate.

关 键 词：说话人识别 语速鲁棒 模型空间分布可视化 MLLR CMLLR 

分 类 号：TN912.34[电子电信—通信与信息系统]