Fast environmental sound classification based on resource adaptive convolutional neural network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Environmental Sound Classification (ESC) is crucial for smart city applications.
• Convolutional Neural Networks (CNNs) improve ESC accuracy but demand high computational resources.
• Existing CNN models are often too complex for embedded systems due to large parameter counts and Floating Point Operations (FLOPs).

Purpose of the Study:

• To develop a lightweight and efficient neural network for environmental sound classification.
• To reduce the computational complexity and hardware requirements of ESC models.
• To enhance the speed and feasibility of deploying ESC on resource-constrained devices.

Main Methods:

• Introduction of a novel Resource Adaptive Convolutional (RAC) module.
• Design of a Resource Adaptive Convolutional Neural Network (RACNN) utilizing the RAC module.
• Integration of RAC modules to upgrade existing CNN architectures for improved efficiency.

Main Results:

• The proposed RACNN model demonstrates superior performance compared to state-of-the-art methods.
• RACNN achieves higher accuracy with significantly lower computational complexity.
• The RAC module efficiently extracts essential time and frequency features from audio data.

Conclusions:

• RACNN offers an effective solution for resource-efficient environmental sound classification.
• The RAC module provides a cost-effective way to generate feature maps, reducing model size and FLOPs.
• This research facilitates the deployment of advanced ESC capabilities in smart city environments and on embedded systems.