基于Cohort相似度的说话人识别

Speaker verification based on Cohort Similarity

本文提出了一种基于Cohort相似度度量的识别方式,训练集外选择出同这个目标说话人比较近似的M个说话人计算M+1维混合高斯,即Cohort模型,,来描述说话人模型,可以很大程度上消除现有系统的不匹配。通过实验,本文提出的基于Cohort的方法,可以将性能提高15%左右,从而证明了该方法的可行性和应用性。

In this paper, a new similarity based on a Cohort is proposed, which will reduce the affection caused by the mismatch of recording environment by choosing M.most similar speakers＇ Models from Cohort as a supplement of the target speaker. At the end of this paper, the experiment showed that the method of Cohort Similarity improved the performance about 15% relatively compared to traditional GMM-UBM systems.