基于SVM评分融合的分类短语音话者确认系统

Classification Speaker Verification System with Short Speech Based on SVM Fusing Scores

对于与文本无关短电话语音(小于30s)的话者确认,在特征参数空间上分类并分别建模的方法,会带来多个子系统输出融合的问题。为了得到最终的评分,同时反映出各个子系统之间的非线性关系以及贡献的不同。本文提出了使用支持向量机(Supportvectormachine,SVM)进行后端评分融合的方法,对输出的两类评分矢量(目标话者和冒认话者)进行分类。在NIST′03数据库上的实验表明,在短语音情况下该方法比评分相加融合方法性能可以相对提高约11%,SVM不仅适用于多子系统的评分级的融合,对其他的多系统多信息的融合也行之有效。

In order to solve the problem of text-independent speaker verification with telephone speech less than 30 s, a method for classifying the speakers′ parameters and separately modeling is presented. However it also brings a problem of the sub-systems fusion. For the sake of getting the last score and reflecting the nonlinear relations and the contributions of each sub-system, a support vector machine (SVM) based fusion model is proposed. The SVM model can classify two kinds of scores (target & non-target). Experiments on the database of NIST′03 show that the verification performance with SVM method can relatively improve about 11% as to score adding method with the short speech. SVM based score combination approach is useful for score fusion of the sub-systems and valuable for other multi-system and multi-information fusions.