基于CGMM-UBM的电话短语音说话人确认  

CGMM-UBM based speaker verification using short telephone speech

作　　者：吴礼福[1] 解焱陆[1] 戴蓓蒨[1] 李辉[1] 

机构地区：[1]中国科学技术大学电子科学与技术系,安徽合肥230026

出　　处：《电路与系统学报》2007年第5期131-136,共6页Journal of Circuits and Systems

基　　金：国家自然科学基金项目(60272039)

摘　　要：针对手机、电话的短语音文本无关说话人确认,本文设计了一种基于分类GMM-UBM(CGMM-UBM)的说话人确认系统。用k-means算法将训练背景模型的语音参数集分类成若干个子空间,并据此进行目标说话人语音数据的子空间分类,再采用GMM-UBM结构为每个子空间分别建立一个子系统,以各个子系统输出评分的线性加权作为系统的输出评分。分类后的模型可以采用较低的混合度,线性加权增强了贡献较大子空间对确认性能的作用。在NIST’03语音库上100个男性话者的实验表明,短语音条件下,分类系统的性能比不分类系统有显著的改进,运算效率较后者也提高很多。This paper proposes a text-independent speaker verification system for short telephone speech-CGMM-UBM. The feature vectors for training UBM are classified into subspaces using k-means algorithm and the feature space of target speaker is classified accordingly. Then the UBM-MAP structure subsystem is constructed for every subspace. The weighted linear sum of scores from every subsystem is computed as the results of the CGMM-UBM. Results of experiments conducted on the NIST＇03 database show that CGMM-UBM system gives better performance than the classical GMM-UBM system, the computation of CGMM-UBM is also lower than the latter.

关 键 词：电话语音 说话人确认 与文本无关 分类高斯混合模型-背景模型 

分 类 号：TN912.34[电子电信—通信与信息系统]