基于GMM的"时间-空间分布模型(TSDM)"及其在说话人辨认中的应用

提出了一种新的注册者模型——“时间一空间分布模型（TSDM）”，传统GMM用参数表征训练矢量的空间分布，但舍弃了训练矢量间的时间联系信息，而TSDM利用基于GMM中均值矢量的高阶协方差矩阵，可向训练矢量的空间分布模型中引入一定程度的训练矢量间时间联系．还给出了TSDM的判据生成方法．实验表明，TSDM能在长训练语句时获得与传统GMM相当的识别性能，在短训练语句时表现得更优秀．

GMM-Based Temporal-Spatial Distribution Model and Its Application to Speaker Identification

A new register model,temporal-spatial distribution model(TSDM),is proposed.Although the traditional GMM provides parameters that can describe the spatial distribution of training vectors,the temporal relation between training vectors is discarded.With high-order covariance generated by the mean vectors provided by GMM,TSDM introduces,to a certain extent,temporal relation between each pair of vectors into spatial distribution.A corresponding score-generated method for TSDM is also proposed.Experiments show that,while providing identification performance similar to GMM with long-duration train contents,TSDM improves identification rate when using short-duration train contents.