Scholarlink

<p>Most music emotion classification methods fail to fully capture the complementary effect between the temporal dynamics and structural dependencies of music features, which limits their ability to model the complex emotional patterns embedded in audio signals. Aiming to address this issue, this paper proposes a hybrid Long Short-Term Memory (LSTM) Networks and Graph Attention Networks (GAT) model for music emotion classification, which effectively combines the temporal modeling capability of LSTM Networks with the relational learning advantage of GAT. The model integrates temporal feature learning and relational modeling to progressively capture emotional representations from low-level acoustic features. First, temporal features are extracted from low-level descriptors (LLD) and Mel frequency cepstrum coefficients (MFCC) using an LSTM network. Then, a two-layer GAT structure creates a dynamic network of relationships between the features. Finally, feature integration is realized through a fully connected layer. The model is evaluated on the Turkish Music Emotion dataset containing 400 songs in four emotion categories. Experimental results show that the proposed hybrid architecture achieves superior classification performance compared to the single-model approach, obtaining an overall accuracy of 96.65%. The method provides an effective approach for improving music emotion recognition by jointly modeling temporal dynamics and relational dependencies.</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

Joint temporal–structural learning for music emotion recognition using LSTM–GAT networks

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Jiale Fang,
  • Bin Guo,
  • Milembolo Miantezila Junior

摘要

Most music emotion classification methods fail to fully capture the complementary effect between the temporal dynamics and structural dependencies of music features, which limits their ability to model the complex emotional patterns embedded in audio signals. Aiming to address this issue, this paper proposes a hybrid Long Short-Term Memory (LSTM) Networks and Graph Attention Networks (GAT) model for music emotion classification, which effectively combines the temporal modeling capability of LSTM Networks with the relational learning advantage of GAT. The model integrates temporal feature learning and relational modeling to progressively capture emotional representations from low-level acoustic features. First, temporal features are extracted from low-level descriptors (LLD) and Mel frequency cepstrum coefficients (MFCC) using an LSTM network. Then, a two-layer GAT structure creates a dynamic network of relationships between the features. Finally, feature integration is realized through a fully connected layer. The model is evaluated on the Turkish Music Emotion dataset containing 400 songs in four emotion categories. Experimental results show that the proposed hybrid architecture achieves superior classification performance compared to the single-model approach, obtaining an overall accuracy of 96.65%. The method provides an effective approach for improving music emotion recognition by jointly modeling temporal dynamics and relational dependencies.