基于核Fisher判别的分类器算法及其在语种识别中的应用研究

Novel Classifier Algorithm Based on Kernel Fisher Discriminant and its Application in Language Recognition

GMM与SVM的建模和识别性能具有较好的互补性,因此GMM-SVM在语种识别中得到广泛使用,以其为基础的GMM-MMI-SVM已成为语种识别的主流研究方法。但是SVM在判别时仅仅使用了训练样本中的一些特殊样本即支持向量,并没有使用全部样本,从而影响了系统识别性能的进一步提高。针对上述问题,提出一种基于核Fisher判别的分类算法——GMM-MMI-KFD。该算法的核心思想是用核Fisher准则(KFD)替代SVM分类准则,从语音片段中提取出特征向量序列,分别通过GMM-MMI分类器与GMM-KFD分类器进行判决打分。相对SVM,KFD更注重语音数据非线性分布的特点,并且将样本向高维空间H上投影后可以最大限度地增大类间距,减小类内距。实验数据表明,GMM-MMI-KFD方法在语种识别中具有更高的识别率。

GMM and SVM have a good complementation on the modeling and recognition performance. Therefore, GMM-MMI-SVM has become a mainstream research method in language recognition. However, SVM only employs some special samples in the training samples,i, e. support vector, but doesn＇t use all samples. This affects further im- provement of system＇s recognition performance. In order to solve this problem, an novel classification algorithm based on Kernel Fisher Discriminant（KFD） was proposed in this paper, called GMM-MMI-KFD. The core idea is the substitu- tion of SVM with KFD, Extracting eigenvector sequence from voice segment, and then inputing them into GMM-MMI and GMM-KFD classifiers respectively, which judge them. Compared to SVM, KFD gets more emphasis on the charac-teristic of nonlinear distribution of voice data. Meanwhile, it can maximize between-class space and minimize within-class space after the projection of samples onto high-dimensional space. The experimental data shows that the GMM-MMI-KFD Classifier has higher recognition rate in language recognition.