基于正交高斯混合模型的说话人识别研究

A Study on Speaker Recognition Using Orthogonal Gaussian Mixture Models

本文介绍了正交高斯混合模型 (OGMM)及其在说话人识别中的具体应用。传统的高斯混合模型 (GMM)常常假定协方差矩阵为对角线矩阵 ,但需大量的混合成员来表征分布情况 ,这将会导致训练量的增加。OGMM的主要思想是在传统的GMM之前先将特征矢量变换到由协方差矩阵的本征向量决定的空间中去 ,这样得到的对角线协方差矩阵可以更准确地反映分布的情况。

This paper introduces the Orthogonal Gaussian mixture model (OGMM) and the application in the speaker recognition Standard GMM assumes diagonal covariance matrices, and needs a large number of mixture components to obtain good approximation which leads to greater training time. This paper proposes a modification to the standard diagonal GMM approach: feature vectors are first transformed to the space spanned by the eigenvectors of the covariance matrix before being applied to the diagonal GMM. An OGMM based speaker recognition experiments show that the performance is better than the standard GMM and has better prospects.