Robust voiceprint recognition with adaptive anti-noise ability based on fitting and compensation  

作　　者：CHEN Zhuang YU Yibiao 

机构地区：[1]School of Electronic Information Engineering,Soochow University Suzhou 215006

出　　处：《Chinese Journal of Acoustics》2022年第3期279-294,共16页声学学报（英文版）

摘　　要：The current voiceprint recognition system has a good performance in a quiet environment,but in the variant noisy background,the performance will decrease sharply due to changes in training and application environment.To solve this problem,this paper proposes a robust noise-adaption voiceprint recognition algorithm,which is i-vector partial least squaresauto encoder(IPLS-AE).IPLS-AE is inspired by noise reduction in i-vector space.The method takes the partial least squares to directly build the relationship between noisy i-vectors and clean i-vectors and then uses auto-encoder to describe the similarity between unknown noises and known noises.Experimental results illustrate that,compared with the typical i-vector maximum a posteriori(IMAP),IPLS-AE has a better compensation performance for various types and different signal-to-noise ratios(SNRs)noises.For the known noise,the relative reduction of equal error rate(EER)and minimum detection cost function(minDCF)are 31.3%and 26.8%,and for the unknown noise,the relative reduction are 28.3%and 25.2%.The results show that the proposed IPLS-AE can effectively compensate for noise,and thereby improve the robustness of the system.

关 键 词：noise sharply performance 

分 类 号：TN912.34[电子电信—通信与信息系统]