Joint spectro-temporal and perceptual feature learning using a dual-track attention network for music genre classification.

Objectives: Music Genre Classification (MGC) plays a pivotal role in information retrieval, underpinning the organization, recommendation, and discovery of music. Current approaches predominantly depend on spectrogram-based convolutional models or handcrafted acoustic features, which inadequately capture the intricate interplay between spectral, temporal, and perceptual cues that define musical genres. Methods: To overcome these limitations, we introduce a dual-track architecture that fuses local spectro-temporal textures with global statistical descriptors through an attention-driven framework. The first track utilizes the Efficient Axis Fusion Residual Attention Network (EAFRAN), incorporating Hybrid Attention Residual Fusion (HARF) and Axis-Integrated Contextual Attention (AICA) modules to model complex time-frequency relationships. The second track employs the Residual Convolutional-Attention Embedding Network (RCAEN), leveraging streaming low-level descriptors and residual multi-head self-attention layers to encode global perceptual features. The complementary feature maps from both tracks are combined to yield an efficient and discriminative representation of musical content. Results: Experiments conducted on the GTZAN benchmark dataset demonstrated that the proposed model attained an accuracy of 98.47 ± 0.6%, outperforming existing state-of-the-art methods. Conclusion: The dual-track attention-based strategy effectively bridges local and global musical features, achieving efficient classification performance and offering a potential foundation for enhanced music information retrieval systems.