基于区分性Model Pushing的语种识别方法

Language recognition based on discriminative Model Pushing

提出一种区分性Model Pushing方法,将SVM训练出的支持向量沿最优分类面的法线方向进行适当移动,增大不同语种间的区分性,然后将移动后的支持向量反向应用于GMM。该方法既保留了SVM的区分性信息,又利用了GMM在短时语音上的优势,同时增加了目标与非目标的区分度。实验结果表明,区分性Model Pushing能有效地提高识别性能。

To improve the performance of short utterances in this model, this work proposes discriminative Model Pushing that moves the support vectors in the direction of the normal to the separation hyperplanes trained by SVM. Then the moved support vectors are pushed back to GMM. By this means, the discriminative information of SVM and the advantage of GMM in short utter- ances recognition are retained, and the diversity between targets and nontargets is enhanced. Experimental results demonstrate that the proposed discriminative Model Pushing produce improvement over the baseline GMM-UBM, GSV-SVM and Model Pushing.