Multi-Resolution Dual-Branch DWT-MobileViT for Ultra-Lightweight Urban Noise Classification

Abstract

Standard environmental sound classification frameworks compress 1D acoustic waveforms into 2D Mel-Spectrograms, discarding raw phase boundaries and flattening fine-grained temporal envelopes. Deep 2D convolutional networks struggle to isolate transient, pulsed mechanical sounds like jackhammers or vehicle horns in noisy municipal environments. They also require excessive parameters and computation for resource constrained deployment. To address these limitations, this paper proposes a lightweight dual-branch hybrid framework combining 1D Discrete Wavelet Transform coefficients with 2D Log-Mel spectrograms, fused via a symmetric Cross-Attention mechanism. The proposed architecture utilizes parameterized Lightweight MobileViT blocks to capture spectral-temporal envelopes while maintaining an extremely compact memory footprint of only 38,322 parameters. The proposed model achieves a 36.50% zeroshot generalization accuracy on ESC-50 across three acoustic datasets, outperforming standard baselines by using significantly fewer parameters. However, this analysis also reports the severe capacity limits and underfitting trade-offs of such extremely small parameter spaces. It demonstrates an expected performance gap with 61.23% accuracy on Urban-Sound8K and 0.2813 mean AUPRC on SONYC-UST. These findings formalize the mathematical boundaries of extreme parameter compression for lightweight acoustic classification.