使用二次特征选择及核融合的语音情感识别

摘要

为提高语音情感识别精度,对基本声学特征构建的多维特征集合,采用二次特征选择方法综合考虑特征参数与情感类别之间的内在特性,从而建立优化的、具有有效情感可分性的特征子集;在语音情感识别阶段,设计二叉树结构的多分类器以综合考虑系统整体性能与复杂度,采用核融合方法改进SVM模型,使用多核SVM识别混淆度最大的情感.算法在Berlin情感语音库五种情感状态的样本上进行验证,实验结果表明二次特征选择与核融合相结合的方法在有效提高情感识别精度的同时,对噪声具有一定的鲁棒性.%To improve the recognition performance of speech emotion recognition, a high dimension acoustic feature set is constructed by basic acoustic features. A secondary feature selection method comprehensively considering the inherent properties between the features and emotions is adopted to select optimal subset with effective emotional recognizability. In the emotion recognition procedure, a binary tree structured multi-class classifier model is adopted to make compromise between total performance and complexity of the system. Kernel fusion method is utilized in SVM model to improve the recognition of the most confusable emotion. The experimental results of five emotions in Berlin database verify the effec-tiveness of the combination of secondary feature selection and kernel fusion on the improvement of emotional recognition accuracies and its robustness on noisy samples.