A lightweight dual branch masking network for environmental sound classification

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Deep neural networks (DNNs) have advanced environmental sound classification (ESC).
• Existing DNNs for ESC often require large models and extensive pretraining, limiting deployment in resource-constrained environments.
• Lightweight models may have limited representational capacity and suboptimal generalization, especially in low-data scenarios.

Purpose of the Study:

• To propose SpectroMaskNet, a compact dual-branch architecture for efficient and robust ESC.
• To address the limitations of existing lightweight models in terms of representational capacity and generalization.
• To enable effective ESC in resource-constrained environments and data-scarce situations.

Main Methods:

• Developed SpectroMaskNet, a compact dual-branch neural network architecture.
• Integrated global-local attention mechanisms to capture long-term temporal dependencies and fine-grained spectral features.
• Employed block-masked spectrogram augmentation to enhance model robustness and generalization, particularly in low-data scenarios.

Main Results:

• SpectroMaskNet achieved high accuracies on benchmark datasets: ESC-10 (97.50%), ESC-50 (95.50%), UrbanSound8K (96.32%), and SpeechCommandV2 (96.52%).
• The model outperformed existing lightweight baselines without requiring large-scale pretraining.
• SpectroMaskNet demonstrated low computational complexity, suitable for efficient and scalable real-world ESC applications.

Conclusions:

• SpectroMaskNet offers a promising solution for efficient and accurate environmental sound classification.
• The proposed architecture effectively balances model compactness with high performance and generalization capabilities.
• The model's efficiency and scalability make it suitable for practical deployment in various ESC applications.