<![CDATA[Efficient environmental sound recognition is essential for smart-city monitoring and edge-based acoustic sensing, yet practical deployment is often constrained by the computational cost and parameter size of deep learning models. This paper presents ShuBAM-TFNet, a lightweight convolutional neural network designed to achieve high recognition accuracy while remaining suitable for resource-limited devices. The proposed architecture processes Log-Mel spectrograms using dilated and time-distributed convolutional layers to capture long-range temporal dependencies and localized temporal patterns without increasing model complexity. A lightweight temporal–frequency attention mechanism is introduced to adaptively emphasize informative regions across time and frequency dimensions, while optional Bottleneck Attention Modules are examined for channel and spatial refinement. The model is evaluated on UrbanSound8K and ESC-10 using standard protocols. On UrbanSound8K, the best-performing variant (without Bottleneck Attention Modules) achieves 91.76 ± 0.90% accuracy with only 70,754 trainable parameters, converging in approximately 26 epochs. Ablation studies confirm that temporal–frequency attention provides the most significant performance gain (≈2–3%), while dilated convolutions accelerate convergence without increasing parameters. Cross-dataset evaluation on ESC-10 achieves 84.00 ± 2.41% accuracy and a macro F1-score of 0.838, demonstrating reasonable generalization. The counterintuitive finding that removing Bottleneck Attention Modules improves performance reveals that channel-spatial attention requires task-specific validation. These results indicate that attention-guided lightweight architectures particularly those emphasizing temporal-frequency modeling over channel-spatial refinement provide a practical alternative for environmental sound recognition on resource-constrained edge devices.]]>
