基于Mahalanobis距离的说话人识别模型研究

Mahalanobis Distance MetricScoring for Speaker Recognition

在传统的i-vector模型中,余弦相似度分类器由于其计算简单和良好的性能被广泛应用。本文根据距离度量学习思想,提出一种新的基于Mahalanobis距离的说话人识别模型。新模型是在i-vector的基础上构建的,语音的i-vector提取出来后,经过白化变换,根据基于信息几何的度量学习算法构造的距离度量矩阵,计算i-vector之间的Mahalanobis距离作为相似度评分依据。实验在NIST2008电话语音库上进行,结果表明基于Mahalanobis距离的识别模型在性能上优于传统的基于余弦相似度的识别模型。

In the i-vector model,cosine similarity scoring method is used for its low complexity and good performance.This paper proposes a new Mahalanobis distance model which adopted the distance metric learning algorithm.The new model is based on the i-vector,and the score between two whitened i-vectors is the Mahalanobis distance.In our paper,the information geometry metric learn algorithm is used to construct the distance metric in the model.The results on NIST 2008 telephone data show that the performance of Mahalanobis distance metric scoring is better than the cosine similarity scoring.