一种新的α-GMM聚类说话人确认算法

A NOVEL α-GMM CLUSTERING APPLIED IN SPEAKER VERIFICATION

针对噪声环境下说话人识别率低的问题,提出一种基于α-GMM聚类和SVM的说话人确认算法。首先计算每位注册话者的α-GMM模型,并计算模型间的α散度,然后以该散度设计聚类算法,对话者的α-GMM模型进行聚类,得到各个类别的聚类中心模型用于训练SVM,进而得到最终识别结果。仿真实验结果表明,该算法相比于传统GMM和SVM具有更高的识别性能和良好的鲁棒性。

A novel speaker verification method based on a-GMM clustering and SVM is proposed in this paper in order to improve the recognition accuracy of speaker in noisy environment. Each of the registered speaker＇s a-GMM is computed first, and the a-divergence between models is computed simultaneously. And then, speaker＇ s a-GMM is clustered using the clustering algorithm designed with this or-divergence. Finally, clustering centre models of every category are derived and are used in training the SVM. Simulative experiment results show that our approach has higher recognition performance and better robustness than the traditional Gaussian mixed model （GMM） and SVM.