Speech Emotion Recognition Based on Fusion Feature Selection Method

JI Xunsheng; LUO Zhan; XIE Jie

doi:10.3969/j.issn.1672-9722.2025.07.017

Abstract：Feature selection plays an important role in reducing model complexity and improving emotion recognition rate for complex and high-dimensional speech emotion features.In order to reduce the feature dimension,a fusion feature selection method is proposed to select the optimal feature subset according to different evaluation criteria through filtering and packaging processes.In ReliefF,the average distance between k nearest neighbor or non-nearest neighbor samples is introduced to estimate the differences between samples,and the weight values of features in the samples are effectively evaluated.The irrelevant feature vectors with nega-tive weight are removed,and the resulting reduced feature vectors are used as the input of the second stage of fusion feature selection.The packaging method uses two-order variation gray wolf optimization algorithm,which considers the interaction between feature and classification algorithm,and introduces two-order variation operator to improve the development efficiency of the algorithm.The weight initialization is completed by random forest method,which retains most significant features in the initial population stage and accelerates the convergence of the algorithm.Experimental results show that this method can effectively reduce the feature dimension and improve the classification performance.Classification accuracy of 90.66%and 78.96%on EMODB and SAVEE datasets is ob-tained using support vector machine classifier,and the feature dimension decreases from 1582 to 366 and 255,respectively.