Speech emotion features selection based on BBO-SVM

Abstract

In order to solve the problem that the emotion feature dimension is too high in speech emotion recognition, an optimization method based on BBO-SVM for invariant set of elements is proposed. BBO-SVM method optimizes the original lower-level speech emotion feature set with higher dimension combining the optimization characteristics of the Biogeography-Based Optimization algorithm (BBO) and the classification training ability of support vector machine (SVM), the purpose is to obtain feature set with low dimension and rich emotional information. This paper first extract 1582-D emotional feature set for each speech file using openSMILE and then randomly divided it into multiple feature subsets, which are used as the original population of BBO algorithm. Corresponds the feature subset to the index of the BBO algorithm. At last, grouping and screening the high redundancy features using BBO. The final results are obtained by multiple iterations, and the cross validation results of support vector machine are used as the criteria for evaluating the generated subsets in the iterative process. The result shows that, for the original speech feature set with high dimension and contains a lot of redundant information, the BBO-SVM method can filter them to obtain feature set with rich emotional information and few redundant components. This provides an optimized feature basis for the establishment of emotion model, which can improve the efficiency of speech recognition process and obtain better recognition results at the same time.