Scholarlink

<p>Speech Emotion Recognition (SER) is vital to human–computer interaction, particularly in healthcare, education, and security. A significant challenge in SER is the high dimensionality of extracted acoustic features, many of which are irrelevant or redundant, degrading classification accuracy. Metaheuristic-based feature selection methods have shown promise, but are susceptible to premature convergence and instability. The present study proposes a novel feature selection technique for SER using the recently developed Walrus Optimizer (WO), a bio-inspired swarm intelligence optimization algorithm based on walrus herd behavior, combined with a K-Nearest Neighbor (KNN) classification model. The WO-KNN strategy aims to identify the optimal subset of informative attributes at a low computational cost. Experiments were conducted on three benchmark datasets: Surrey Audio-Visual Expressed Emotion (SAVEE), Ryerson Audio-Visual Database (RAVDESS), and Arabic Emirati-accented speech. The proposed WO-KNN methodology was compared with well-known metaheuristic feature selection techniques. Performance analysis demonstrates that WO-KNN consistently achieves better classification performance, faster convergence, and smaller feature subsets across datasets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Walrus optimizer-based feature selection for robust speech emotion recognition

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Tingting Fu

摘要

Speech Emotion Recognition (SER) is vital to human–computer interaction, particularly in healthcare, education, and security. A significant challenge in SER is the high dimensionality of extracted acoustic features, many of which are irrelevant or redundant, degrading classification accuracy. Metaheuristic-based feature selection methods have shown promise, but are susceptible to premature convergence and instability. The present study proposes a novel feature selection technique for SER using the recently developed Walrus Optimizer (WO), a bio-inspired swarm intelligence optimization algorithm based on walrus herd behavior, combined with a K-Nearest Neighbor (KNN) classification model. The WO-KNN strategy aims to identify the optimal subset of informative attributes at a low computational cost. Experiments were conducted on three benchmark datasets: Surrey Audio-Visual Expressed Emotion (SAVEE), Ryerson Audio-Visual Database (RAVDESS), and Arabic Emirati-accented speech. The proposed WO-KNN methodology was compared with well-known metaheuristic feature selection techniques. Performance analysis demonstrates that WO-KNN consistently achieves better classification performance, faster convergence, and smaller feature subsets across datasets.