Scholarlink

<p>Arabic speech emotion recognition faces significant challenges due to limited annotated datasets and the linguistic complexity of Arabic dialects. This research presents GAN-AraEmo, a novel generative adversarial network framework specifically designed for augmenting low-resource Arabic emotional speech datasets. The proposed architecture incorporates a multi-scale generator with attention mechanisms and an enhanced discriminator utilizing spectral normalization to ensure high-quality synthetic data generation. A novel emotional consistency loss function maintains emotional authenticity while generating diverse speech patterns across multiple Arabic dialects. The framework addresses data scarcity by generating synthetic emotional speech samples that preserve both linguistic authenticity and emotional characteristics. Experimental validation on three Arabic emotional speech datasets demonstrates substantial performance improvements, with GAN-AraEmo achieving 91.2% recognition accuracy compared to 67.5% baseline performance, representing a 23.7% improvement. Generated synthetic samples exhibit high perceptual quality with 4.2/5.0 mean opinion score and maintain 94.3% emotional authenticity in human evaluation. Cross-dialect evaluation reveals robust generalization with only 9.8% performance degradation across dialectal boundaries compared to 20.1% for baseline methods. The research contributes the first specialized GAN architecture for Arabic SER data augmentation, comprehensive evaluation across Arabic dialects, and significant advancement toward culturally aware emotion recognition systems for Arabic-speaking communities.</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

GAN-AraEmo: Generative Data Augmentation for Low-Resource Arabic Speech Emotion Recognition

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Hend Alshaya,
  • Wided Bouchelligua

摘要

Arabic speech emotion recognition faces significant challenges due to limited annotated datasets and the linguistic complexity of Arabic dialects. This research presents GAN-AraEmo, a novel generative adversarial network framework specifically designed for augmenting low-resource Arabic emotional speech datasets. The proposed architecture incorporates a multi-scale generator with attention mechanisms and an enhanced discriminator utilizing spectral normalization to ensure high-quality synthetic data generation. A novel emotional consistency loss function maintains emotional authenticity while generating diverse speech patterns across multiple Arabic dialects. The framework addresses data scarcity by generating synthetic emotional speech samples that preserve both linguistic authenticity and emotional characteristics. Experimental validation on three Arabic emotional speech datasets demonstrates substantial performance improvements, with GAN-AraEmo achieving 91.2% recognition accuracy compared to 67.5% baseline performance, representing a 23.7% improvement. Generated synthetic samples exhibit high perceptual quality with 4.2/5.0 mean opinion score and maintain 94.3% emotional authenticity in human evaluation. Cross-dialect evaluation reveals robust generalization with only 9.8% performance degradation across dialectal boundaries compared to 20.1% for baseline methods. The research contributes the first specialized GAN architecture for Arabic SER data augmentation, comprehensive evaluation across Arabic dialects, and significant advancement toward culturally aware emotion recognition systems for Arabic-speaking communities.