基于一种新的特征提取的语音辨识

Novel Feature Extraction Method for Speaker Identification

提出了一种新的二次特征提取的方法应用于说话人语音辨识。首先,通过基于熵的特征筛选方法,有效地剔除不重要或者噪声特征,消除语音特征的冗余,并获得其重要性排序,减少语音特征矢量的维数。然后,采用Fisher准则进一步进行参数选择,按Fisher比的大小选择特征向量作为投影轴,将高维空间中的特征矢量映射到低维的特征判别空间,然后以SVM作为分类器实现说话人辨识系统。实验结果表明,本文提出的方法在不影响识别率的情况下可以对输入数据有效降维,在噪音环境下取得了较好的识别效果,增加了系统的鲁棒性。

A novel method based on combining a secondary feature extraction method for speaker identification is proposed. Firstly, the entropy-based feature selection approach is exploited to reduce the dimension of the input vectors ,by which an important feature subset can be obtained, whilst the redundancy and relevance of the voice attributes can be eliminated effectively. Secondly, fisher discriminant criterion is used for parametric selection. The feature vectors are chosen according to their fisherby ratio and are used as projection vectors. Then the feature vectors in High-dimensional space are mappedto low-dimensional determination space. Finally, SVM is used as the classifier to complement the recognition system. The experimental results show that the storge can be reduced remarkably without deteriorating the recognition performance by the proposed method compared with other reduced algorithms. Whilst, it increases the robustness of the identification system.