似然得分归一化及其在与文本无关说话人确认中的应用  

Likelihood Score Normalization and Its Application in Text-Independent Speaker Verification

作　　者：邓浩江[1] 杜利民[1] 万洪杰[1] 

机构地区：[1]中国科学院声学研究所语音交互技术研究中心,北京100080

出　　处：《电子与信息学报》2005年第7期1025-1029,共5页Journal of Electronics & Information Technology

摘　　要：该文研究了似然得分归一化方法的原理,建立了基于自适应GMM模型的说话人确认系统,并将非特定人的背景模型与特定人的cohort模型相结合,提出了混合归一化的方法。在电话语音条件下,该文比较了不同得分归一化方法对确认系统性能的影响。实验表明,在自适应GMM模型似然比得分的基础上,T-cohort与通用背景模型混合归一化能获得最佳识别效果。当错误拒绝率为5%时,该方法可以获得0.5%的错误接受率,远远低于采用通用背景模型归一化方法的2%。In this paper, the methodology of likelihood score normalization is studied. The text-independent speaker recognition system based on the adapted Gaussion Mixture Models(GMMs) is established, and the approach to normalize scores combining speaker-independent background model and the speaker-dependent models of cohort speaker sets are proposed. The speaker verification experiments over telephone channels show that based on the likelihood ratio of adapted GMMs system, both cohort normalization and hybrid score normalization approaches can improve the verification performance of baseline system using Universal Background Model (UBM). Specially, the hybrid approach combining UBM and cohort models selected during testing (T-cohort normalization) achieve the best performance. At a miss probability of 5%, the hybrid approach using UBM and T-cohort models reduce the false alarm rate to 0.5% compared to 2% for the baseline.

关 键 词：说话人确认 高斯混合模型 得分归一化 与文本无关 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]