摘要
在文本无关的说话人识别中,采用均值超向量作为特征向量的支持向量机系统性能已经超过了传统的混合高斯-通用背景模型系统,但是信道的影响在均值超向量上仍然存在。该文对因子分析算法进行修改后,可以解决均值超向量的信道问题,能够取得优于扰动属性映射的性能,更重要的是采用因子分析的系统的稳定性可以得到保证。在NIST 2006说话人测试数据库上,利用该文的方法能够取得等错误率6.0%。
In the text-independent speaker recognition system, the mean-supervector of Gaussian Mixture Models (GMM) and Support Vector Machine (SVM) system can outperform the traditional GMM and Universal Background Models (UBM) system, but the session variability is still one of the most important reasons that deteriorate the performance. In this paper, the factor analysis is tailored to solve the session variability problem of GMM mean-supervector. The proposed algorithm can outperform the Nuisance Attribute Projection (NAP) algorithm. Furthermore, the proposed system based on factor analysis is more stable than the system based on NAP. In the NIST 2006 SRE corpus, the Equal Error Rate (EER) of the proposed system can obtain 6.0%.
出处
《电子与信息学报》
EI
CSCD
北大核心
2009年第2期302-305,共4页
Journal of Electronics & Information Technology
关键词
说话人确认
超向量
联合因子分析
扰动属性映射
Speaker verification
Supervector
Joint factor analysis
Nuisance Attribute Projection(NAP)