Scholarlink

<p>Music Emotion Recognition (MER) is a fundamental technology for applications ranging from personalized recommendation to affective computing. However, effectively fusing audio and symbolic data remains challenging due to discrepancies in temporal granularity and the high computational cost of modeling long-term dependencies. To address these issues, this paper proposes ATEM-BiAlign, a lightweight bimodal framework that integrates modality-specific encoding, collaborative alignment, and gated fusion. Specifically, the audio branch employs a Conformer-based encoder with Linear Relative Attention to efficiently capture continuous emotional variations, while the symbolic branch utilizes a Segment-BoE strategy to extract robust rhythmic statistics. A bidirectional collaborative attention mechanism with token gating is further designed to align these heterogeneous modalities. Experiments on the EMOPIA and VGMIDI datasets demonstrate that our method achieves state-of-the-art performance, with an accuracy of 0.84 on EMOPIA, significantly outperforming strong baselines while maintaining low computational complexity.</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

ATEM-BiAlign:Audio-MIDI Bimodal Music Emotion Recognition via Linear Attention

  • 期刊论文

  • 发表日期：

    2026年3月22日

查看全文
  • Xun Liu,
  • Jing Guo,
  • Guangye Li

摘要

Music Emotion Recognition (MER) is a fundamental technology for applications ranging from personalized recommendation to affective computing. However, effectively fusing audio and symbolic data remains challenging due to discrepancies in temporal granularity and the high computational cost of modeling long-term dependencies. To address these issues, this paper proposes ATEM-BiAlign, a lightweight bimodal framework that integrates modality-specific encoding, collaborative alignment, and gated fusion. Specifically, the audio branch employs a Conformer-based encoder with Linear Relative Attention to efficiently capture continuous emotional variations, while the symbolic branch utilizes a Segment-BoE strategy to extract robust rhythmic statistics. A bidirectional collaborative attention mechanism with token gating is further designed to align these heterogeneous modalities. Experiments on the EMOPIA and VGMIDI datasets demonstrate that our method achieves state-of-the-art performance, with an accuracy of 0.84 on EMOPIA, significantly outperforming strong baselines while maintaining low computational complexity.