Music Information Classification and Chord Recognition Based on Dual-Modal Fusion and Contextual Information Extraction

To address the bottleneck of traditional music information classification (MIC) methods in balancing local textures and global context, as well as insufficient fusion of multimodal features, a unified analysis framework based on innovative combinations of existing computational intelligence components is proposed. In terms of methodology, the study innovatively integrates four key modules, including bimodal fusion of Mel spectrogram and low-level temporal features, a sliding window multi-head attention mechanism (SW-MSA) for capturing multi-scale information, a position convolutional layer, and a dynamic gating mechanism. The results show that in the music genre classification (MGC) task, the proposed network achieves an average F1-score of 96.13% on the noisy and imbalanced “MagnaTagATune” (MTAT) dataset, and a stable final training accuracy of 98.75% on the clean and balanced “Free Music Archive-small” (FMA-small) dataset, outperforming the baseline model. In the automatic chord recognition (ACR) task, the proposed network achieves a Tetrads class weighted chord symbol recall score of 92.23% on the Isophonics dataset. When dealing with 170 types of unbalanced chords, this method improves the recognition accuracy of low-frequency chords by 7.29% compared to the Chord Transformer model. This indicates that the proposed model can effectively improve classification and recognition performance. This study provides an innovative model architecture and a feasible technical path for high-precision automatic analysis of music information.