ESC-NAS: Environment Sound Classification Using Hardware-Aware Neural Architecture Search for the Edge

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Deep learning and IoT integration are crucial for smart solutions, enabling real-time offline operations with enhanced accuracy and reduced resource needs.
• Environmental sound classification (ESC) on edge devices faces challenges due to limited computational resources and the need for processing raw audio data.

Purpose of the Study:

• To propose ESC-NAS, a novel hardware-aware neural architecture search (NAS) approach for designing deep convolutional neural networks (CNNs) for ESC applications.
• To develop CNN architectures optimized for raw audio input, focusing on minimizing resource consumption while maintaining high accuracy for edge deployment.

Main Methods:

• Developed a cell-based NAS search space incorporating 2D convolution, batch normalization, and max pooling layers for raw audio feature extraction.
• Employed a black-box Bayesian optimization search strategy to explore the NAS space.
• Evaluated model architectures using hardware simulation to assess performance and resource consumption.

Main Results:

• ESC-NAS achieved an optimal trade-off between model performance and resource consumption compared to existing methods.
• Achieved high accuracies on benchmark datasets: 85.78% (FSC22), 81.25% (UrbanSound8K), 96.25% (ESC-10), and 81.0% (ESC-50).
• Generated models with optimal sizes and parameter counts suitable for edge deployment.

Conclusions:

• ESC-NAS effectively designs efficient deep learning models for environmental sound classification on resource-constrained edge devices.
• The hardware-aware NAS approach enables the creation of specialized CNNs that excel in processing raw audio data.
• The developed models offer a practical solution for real-world applications requiring accurate and efficient sound analysis at the edge.