Scholarlink

Speech emotion recognition (SER) aims to infer human emotions based on speech information. Effectively and comprehensively combining the various pieces of information in the audio for accurate judgment remains a significant challenge. In this paper, we propose a novel SER network based on a Collaborative Enhanced Attention (CEA) mechanism. Initially, we apply wavelet transform to translate the raw speech into the frequency domain, obtaining log spectrograms. We then use EfficientNet and Mamba to extract acoustic features. Finally, CEA and Co-attention are used to fuse these features with high-level acoustic information embedded through the WavLM model. The CEA is specifically designed to refine the acoustic information within WavLM features by accentuating different aspects of the signal, thereby enhancing the feature representation from multiple perspectives. We evaluate our method on the IEMOCAP dataset, where the weighted accuracy and unweighted accuracy improved by 4.96% and 4.71%, respectively. Our code is available on https://github.com/Chara2001/CEA-SER.git .

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

Collaborative Enhanced Attention for Speech Emotion Recognition Based on Multimodal Acoustic Information Fusion

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ping Lin,
  • Can Bu,
  • Yongming Huang

摘要

Speech emotion recognition (SER) aims to infer human emotions based on speech information. Effectively and comprehensively combining the various pieces of information in the audio for accurate judgment remains a significant challenge. In this paper, we propose a novel SER network based on a Collaborative Enhanced Attention (CEA) mechanism. Initially, we apply wavelet transform to translate the raw speech into the frequency domain, obtaining log spectrograms. We then use EfficientNet and Mamba to extract acoustic features. Finally, CEA and Co-attention are used to fuse these features with high-level acoustic information embedded through the WavLM model. The CEA is specifically designed to refine the acoustic information within WavLM features by accentuating different aspects of the signal, thereby enhancing the feature representation from multiple perspectives. We evaluate our method on the IEMOCAP dataset, where the weighted accuracy and unweighted accuracy improved by 4.96% and 4.71%, respectively. Our code is available on https://github.com/Chara2001/CEA-SER.git .