Attention-Assisted Multi-Scale Music Spectral Representation Learning for Intelligent Emotion Quantification

Abstract Accurate quantification of musical emotions remains challenging due to the complex relationship between acoustic structures and human affective perception. Existing approaches mainly focus on fixed spectral representations or categorical emotion recognition, which limits their ability to capture hierarchical frequency-temporal patterns and continuous emotional variations in music signals. This paper proposes MASAN-Net, an attention-assisted multi-scale spectral representation learning framework for continuous music emotion quantification. The proposed framework employs multi-scale convolutional feature extraction to learn hierarchical spectral representations and introduces frequency-temporal attention mechanisms to adaptively identify emotion-related acoustic patterns. A regression-based prediction module is further developed to estimate continuous Valence and Arousal dimensions from music spectrograms. Extensive experiments are conducted on the DEAM dataset, with cross-dataset validation on the MTG-Jamendo dataset to evaluate generalization capability. Compared with traditional machine learning methods and advanced deep learning baselines, MASAN-Net achieves superior performance under multiple regression metrics, including MSE, RMSE, MAE, and Pearson correlation. Ablation studies and repeated experiments further verify the effectiveness and stability of the proposed framework. Attention visualization provides interpretable insights into the spectral-temporal regions associated with emotional prediction. The proposed method offers an effective and explainable solution for intelligent music emotion analysis.