联合因子分析和稀疏表示在稳健性说话人确认中的应用  

Robust speaker verification using sparse representation on joint factor analysis

作　　者：杨海[1] 张翔[1] 梁春燕[1] 索宏彬[1] 颜永红[1] 

机构地区：[1]中国科学院声学研究所中国科学院语言声学与内容理解重点实验室,北京100190

出　　处：《声学学报》2012年第5期548-552,共5页Acta Acustica

基　　金：国家科技支撑计划(2008BAI50B03);国家自然科学基金(10925419,90920302,10874203,60875014)经费资助

摘　　要：在说话人确认的任务中,为了解决信道失配问题,提高系统性能,引入了联合因子分析和稀疏表示算法。首先利用联合因子分析算法去除信道干扰,得到与信道无关的说话人因子,然后在稀疏表示算法中利用说话人因子构建过完备字典,求解稀疏最优化问题计算说话人得分。由于此方法有机结合了联合因子分析算法的信道鲁棒性和稀疏表示的鉴别性,使用此算法构建的系统在NIST SRE 2008电话训练、电话测试数据集上性能表现良好,相对于联合因子分析-支持向量机系统在性能上有竞争性,在原理上有互异性,系统融合更带来了最小检测代价指标上4.91%的性能提升。实验表明使用联合因子分析与稀疏表示进行说话人确认是可行的。This paper introduced sparse representation on joint factor analysis to solve the channel mismatch problem and to improve system performance. This algorithm uses joint factor analysis to generate the speaker factors space and construct the over-complete dictionary to calculate speaker score by solving the optimization problem. The minimum detection cost function （minDCF） of the system with sparse representation on joint factor analysis gave good performance on NIST speaker recognition evaluation （SRE） 2008 telephone to telephone test corpus. Because the sparse representation algorithm and the support vector machine classification algorithm also have a good complementary, the fusion of JFA-SR and JFA-SVM can achieve 4.91% reduction in minDCF. The results of the experiments show that speaker verification using sparse representation on joint factor analysis is feasible and has a great future.

关 键 词：因子分析 稀疏表示 稳健性 说话人确认 信道干扰 应用 最优化问题 支持向量机 

分 类 号：TN912.34[电子电信—通信与信息系统]