EMD-Tnorm得分规整策略在说话人确认中的应用

A new score normalizaion algorithm based on EMD-Tnorm for speaker verification

从两个方面对确认系统进行了改进,在模型方面,扩展了MixMax模型,对复杂的背景噪声等干扰因素在训练说话人模型的同时也进行了建模,最大程度上消除噪声的影响,对说话人的特征分布进行了更真实的表征;在得分方面,提出了一种改进的得分规整策略,基于EMD距离从所有背景说话人集合中自适应选择最接近的一定数量的模型构成说话人特定的背景集合,从而进行得分归一化。实验结果表明,该方法能够同时针对说话人和测试环境的不同进行补偿,进一步降低了误识率和漏警率,获得了很好的确认性能。

In this paper,the verification system from two aspects was improved.On one hand,we extended MixMax model that the EMD（earth mover＇s distance） can be applied,which can remove the disturbance of noise;on the other hand,we improved the Tnorm score normalization method based on the EMD.Experimental results show that this method can compensate the speaker-dependent and test-dependent variability,also show a stable performance improvement by decreasing the FA and FR.