Music genre classification through federated deep harmonic convolution and attention learning

Abstract

Intelligent music recommendation and retrieval systems need genre categorization, yet class imbalance, overlapping genre features, decentralized data privacy, and restricted deployment efficiency continue to challenge present methods. AAI-HarmoCNN-AttnNet, a privacy-conscious federated deep learning architecture for accurate and scalable music genre categorisation, addresses these concerns. The proposed model captures fine-grained spectrum cues and long-range temporal relationships using harmonic-sensitive convolutional layers and dual-path attention. Federated learning allows dispersed clients to optimize while preserving raw audio data. A hybrid hyperparameter optimization technique combining Egret Swarm Optimization and Golden Jackal Optimization improves convergence stability and generalization. AAIHarmoCNN-AttnNet outperforms thirteen competitive baselines, including CRNN, Bi-GRU with attention, and recent self-supervised methods, with 99.1% classification accuracy, 98.9% precision, 98.8% recall, and 97.4% Genre Diversity Sensitivity (GDS) score, in extensive NCASI benchmark dataset experiments. Federated evaluations show robust convergence under non-IID client distributions, with tightly clustered client-wise accuracy above 99% and decreased inter-client variation. Ablation experiments confirm the complimentary contributions of harmonic convolution and dual-path attention, while ROC analysis shows excellent discrimination with high true-positive rates and low false-positive rates. For real-time deployment, edge device resource profiling has low inference latency, small model size, and balanced power efficiency. This shows that AAI-HarmoCNN-AttnNet is a strong, privacy-preserving, and deployment-ready solution for federated music genre classification in current intelligent audio systems.