Scholarlink

<p>The presented paper presents a new method of music emotion recognition (MER) based on a multi-scale convolutional neural network with attention (MS-ACNN). The research will categorize music into two general emotional conditions: happy and sad. The proposed solution addresses the main issues of MER by possessing a powerful preprocessing pipeline that involves the elimination of noise, amplitude normalization, and segmentation, thereby uniformizing the data for appropriate classification. The primary feature representation is the log-Mel spectrograms, which are both spectral and temporal features that are represented in a perceptually significant manner. It is enhanced with pitch shifting, spec augmentation, and time stretching to improve the model and minimize overfitting. The MS-ACNN architecture is designed to detect multiscale features, encompassing both fine-grained and global temporal-spectral music trends. It features an attention system that focuses on aspects of the audio associated with emotions, thereby increasing the accuracy of classification. Accuracy, precision, recall, and F1-score are quality measurements of model performance, demonstrating that the model has improved considerably compared to traditional MER methods. Additionally, attention heatmaps show that the model effectively identifies key areas in the music that aid in emotion recognition. This is a promising method of recognizing emotions in real-time in music-based applications.</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

Research on music emotion recognition and classification model based on deep learning

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Pang Jinshan

摘要

The presented paper presents a new method of music emotion recognition (MER) based on a multi-scale convolutional neural network with attention (MS-ACNN). The research will categorize music into two general emotional conditions: happy and sad. The proposed solution addresses the main issues of MER by possessing a powerful preprocessing pipeline that involves the elimination of noise, amplitude normalization, and segmentation, thereby uniformizing the data for appropriate classification. The primary feature representation is the log-Mel spectrograms, which are both spectral and temporal features that are represented in a perceptually significant manner. It is enhanced with pitch shifting, spec augmentation, and time stretching to improve the model and minimize overfitting. The MS-ACNN architecture is designed to detect multiscale features, encompassing both fine-grained and global temporal-spectral music trends. It features an attention system that focuses on aspects of the audio associated with emotions, thereby increasing the accuracy of classification. Accuracy, precision, recall, and F1-score are quality measurements of model performance, demonstrating that the model has improved considerably compared to traditional MER methods. Additionally, attention heatmaps show that the model effectively identifies key areas in the music that aid in emotion recognition. This is a promising method of recognizing emotions in real-time in music-based applications.