噪声自适应拟合补偿的鲁棒性声纹识别算法  

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

作　　者：陈壮 俞一彪[1] CHEN Zhuang;YU Yibiao(School of Electronic Information Engineering,Soochow University,Suzhou 215006)

机构地区：[1]苏州大学电子信息学院,苏州215006

出　　处：《声学学报》2022年第1期151-160,共10页Acta Acustica

摘　　要：目前的声纹识别系统在安静环境下性能已经很高,但处在复杂噪声背景时,由于训练和应用环境的不同会导致系统性能急剧下降。为解决这一问题,从i-vector空间降噪思想出发,通过偏最小二乘算法直接推导含噪i-vector与纯净i-vector之间的关系,并采用自编码器衡量未知噪声类型的含噪i-vector与已知噪声类型的含噪i-vector之间的相似性,提出一种i-vector空间噪声自适应的鲁棒性声纹识别算法(IPLS-AE)。实验结果显示,相对典型的最大后验概率估计算法(IMAP),提出的IPLS-AE算法对各信噪比与各类型的噪声都有较好的补偿性能,等错误率和最小检测代价函数对已知噪声最高有31.3%和26.8%的相对改善,对未知噪声最高有28.3%和25.2%的相对改善。结果说明,IPLS-AE算法能可靠地识别并补偿噪声,提升系统的鲁棒性。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,starting from the motivation of noise reduction in i-vector space,this paper proposes a robust noise adaption voiceprint recognition algorithm,which is i-vector partial least squares-auto encoder(IPLS-AE).IPLS-AE takes the partial least squares method to directly build the relationship between noisy i-vectors and clean i-vectors and then uses auto-encoder to describe the similarity between unknown and known noises.Experimental results illustrate that,compared to the typical i-vector maximum a posteriori(IMAP),IPLS-AE has a better compensation performance for various noises which are different types and signal-to-noise ratios(SNRs).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：声纹识别 偏最小二乘算法 自编码器 等错误率 直接推导 最大后验概率估计 IMAP 未知噪声 

分 类 号：TN912.34[电子电信—通信与信息系统]