Graph-Based Audio Classification Using Pre-Trained Models and Graph Neural Networks.

Area of Science:

• Acoustics and Signal Processing
• Machine Learning
• Environmental Science

Background:

• Sound classification is vital for acoustic data analysis and environmental monitoring.
• Traditional methods may not fully capture complex audio patterns.
• Graph representation offers a novel approach to audio data.

Purpose of the Study:

• To explore audio data representation as graphs for sound classification.
• To evaluate the performance of various graph neural networks (GNNs) on audio tasks.
• To identify the most effective GNN model for environmental sound analysis.

Main Methods:

• Utilized pre-trained audio models to extract deep audio features.
• Constructed graphs using extracted features as node information.
• Trained and compared graph convolutional networks (GCNs), GraphSAGE, and graph attention networks (GATs).

Main Results:

• Graph representation of audio data proved effective for classification.
• GNNs demonstrated competitive performance in sound classification tasks.
• The graph attention network (GAT) model achieved the highest accuracy: 83% for environmental sounds and 91% for land cover identification.

Conclusions:

• Graph representation learning is a promising technique for audio data analysis.
• GNNs, particularly GATs, offer a powerful tool for multi-class audio classification.
• This methodology enhances the interpretation and application of acoustic data in environmental contexts.