基于聚类和排序高斯混合模型的说话人确认

Speaker Validation Based on Clustering and Sorted Gaussian Mixture Model

基于高斯混合模型(GMM)-通用背景模型(UBM)结构的说话人确认系统不能完全表现说话人的个性特征信息。为此,将聚类方法和排序高斯混合模型相结合,对每个高斯分量按照对应排序值顺序排列,并对UBM进行训练。基于NIST 06 8side-1side数据库的实验结果表明,该方法能在基本保持系统识别性能的前提下,降低UBM的训练运算量。

Gaussian Mixture Model（GMM）——universal background model is used for most of text-independent speaker validation systems in the past decade.This paper proposes a new structure of GMM——Sorted Gaussian Mixture Model,in which each Gaussian components in the universal background model are arranged in corresponding value order,it is an approach to combine with the clustering method to train UBM.Experiments on the 2006 NIST 8side-1side subset speaker recognition evaluation task show that after using this approach,the amount of calculation can be reduced,and under certain search width conditions,almost no reduction in recognition performance.