Abstract: For the F-score feature selection algorithm can not reveal the mutual information among features, the method of removing the redundancy is applied to improve the F-score algorithm. Using the German emotional speech database EMO-DB, based on the extraction of speech emotion features, the paper uses the classification accuracy of SVM to choose the best feature subset. Compared with the F-score method, the improved feature selection algorithm can achieve dimension reduction substantially, select an effective feature subset, and obtain an ideal speech emotion recognition accuracy.

Key words: feature selection, F-score, mutual information, Support Vector Machine（SVM）, speech emotion recognition

摘要： 针对F-score特征选择算法不能揭示特征间互信息而不能有效降维这一问题，应用去相关的方法对F-score进行改进，利用德语情感语音库EMO-DB，在提取语音情感特征的基础上，根据支持向量机（SVM）的分类精度选择出分类效果最佳的特征子集。与F-score特征选择算法对比，改进后的算法实现了候选特征集较大幅度的降维，选择出了有效的特征子集，同时得到了较理想的语音情感识别效果。

关键词: 特征选择, F-score, 互信息, 支持向量机, 语音情感识别