EdgeCAAI-Net: Artist-Invariant Network for On-Device Music Genre Classification
Year: 2026
Type: conference-paper
Abstract: Music genre classification has been studied extensively in music information retrieval (MIR), but practical edge deployment is still difficult because high-performing audio models are usually large. A second issue is evaluation bias: in standard track-disjoint splits, tracks from the same artist can appear in both training and test sets, leading to artist leakage. In this work, we present EdgeCAAI-Net (Edge Compute-Adaptive, Artist-Invariant Network), a 1.42M-parameter model built on Conformer-Lite blocks. The architecture combines (1) multi-exit inference for confidence-based early stopping, (2) a budget-aware loss to encourage earlier exits, and (3) a gradient reversal branch to reduce artist-specific cues in the shared representation. On FMA-Small, moving from track-disjoint to artist-disjoint evaluation lowers macro-F1 by 0.105 (18.8% relative) across architectures, clearly indicating leakage. Ablation studies show that early exits and budget-aware training improve accuracy, while GRL introduces a modest trade-off (−0.012 F1) in return for improved artist invariance. Overall, EdgeCAAI-Net maintains competitive accuracy with a model size suitable for real-time mobile and IoT use.
Institution: Vishwakarma Institute of Technology
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166804722