基于可区分性加权的模糊核说话人识别  

A Fuzzy Kernel with Discriminative Weighted Method for Speaker Recognition

作　　者：林琳[1] 王树勋[1] 陈建[1] 

机构地区：[1]吉林大学通信工程学院,吉林长春130022

出　　处：《电子学报》2008年第7期1446-1450,共5页Acta Electronica Sinica

摘　　要：针对训练和识别语音数据较少的情况,本文提出了一种新的说话人识别算法.通过核映射,在高维特征空间对说话人的语音特征进行模糊矢量量化.为了增加说话人之间的可区分性,提出了一种基于高维特征空间的码字矢量的权值分配方法,对具有较强区分性的码字矢量分配较大的权值,并将产生的权值和说话人的码书一起形成说话人数据库.识别时,提出一种模糊核加权最近邻近分类器,在高维特征空间中对说话人进行匹配.实验表明,该算法在训练语音少于8s,识别语音为1s时,能够得到较好的识别结果.As to small amounts of training and test speech data, it proposed a new speaker recognition algorithm. By the kernel mapping,it used the fuzzy vector quantization to quantize the speakers＇ speech features in the high dimensional feature space.in order to improve the discriminations of different speakers, it presented a novel weights assignment method. It assigned the lager weight to the code vector with higher discriminative power. Then, it used the codebooks and weights to form the speakers＇ database. In the matching phase, it proposed a fuzzy kernel weighted nearest prototype classifier, which can identify different speakers in the high dimensional space.Experimental results show that when the Iraining speech data is less than 8s,and test speech data is 1 s, this algorithm can get good performance.

关 键 词：说话人识别 少量语音数据 可区分性权值 模糊核加权最近邻近分类器 模糊核矢量量化 

分 类 号：TN912.3[电子电信—通信与信息系统]