摘要
针对话者识别系统中特征向量不定长和交叉信道干扰等问题,提出一种基于超向量的扰动属性投影(NAP)核函数。该函数是一种新型的序列核函数,使支持向量机能在整体语音序列上分类,移除核函数空间中与话者识别无关的信道子空间信息。仿真实验结果表明,该函数可有效提高支持向量机的分类性能和话者识别系统的识别准确率。
For the sake of solving the problem of variable-length feature vectors and channel impact which existed in speaker verification,a novel kernel function based on Gaussian Mixture Model（GMM） supervector,called Nuisance Attribute Projection（NAP） mapping KL divergence linear kernel function,is proposed in this paper.This function can not only be in the interest of enabling Support Vector Machine（SVM） to classify on whole audio sequences,but also has the benefit that channel subspace,which causes variability,is removed in kernel space.By doing so,the classification performances of SVM and verification accuracy of system are improved excellently.Simulation experimental results demonstrate the effectiveness of this kernel function.
出处
《计算机工程》
CAS
CSCD
北大核心
2010年第8期194-196,共3页
Computer Engineering
关键词
扰动属性投影
高斯混合模型超向量
话者识别
支持向量机
Nuisance Attribute Projection（NAP）
Gaussian Mixture Model（GMM） supervector
speaker verification
Support Vector Machine（SVM）
