An intelligent cloud-based architecture for music classification and recursive metadata management in digital libraries
Huifang Qi, Hao Jiang · Australian Journal of Electrical & Electronics Engineering · 2025
The rapid growth of digital music collections has made accurate metadata classification essential for effective organisation and retrieval. This study presents a cloud-based architecture for music genre recognition that leverages recent Deep Learning advances through a Frequency-Aware Transformer with Learnable Adaptive Positional Encoding (LAPE), Frequency-Adaptive Multi-Head Self-Attention (FA-MHSA), and layer-wise residual gating. Audio inputs undergo Spectral Flux-Based Dynamic Segmentation, amplitude normalisation, and silence trimming, while hierarchical features are extracted via multi-resolution Mel-Frequency Cepstral Coefficients (MFCCs) and Mel-spectrograms, combined through a Learnable Weighted Fusion process. Experiments on the GTZAN dataset demonstrate exceptional performance, achieving 99% accuracy, 98.98% precision, 99.07% recall, and 99.02% F1-score—surpassing existing methods. The model maintained an average accuracy of 99% across ten genres, with class-wise accuracy ranging from 97% to 100%, indicating strong generalisation without bias. Further validation through confusion matrices, ROC curves, and training dynamics confirms the model’s capability to capture complex time–frequency dependencies in music. Overall, the proposed system provides a highly accurate, robust, and scalable solution for real-time music genre classification and metadata management in large-scale digital music libraries.