Fast grid-free strength mapping of multiple sound sources from microphone array data using a Transformer architecture

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Traditional microphone array methods for sound source characterization often rely on focus-grids, leading to high computational costs or limited accuracy.
• Existing grid-free methods may require multiple models for varying numbers of sound sources.

Purpose of the Study:

• To develop a deep learning-based, grid-free method for accurate sound source characterization.
• To enable a single model to identify an unknown number of sound sources simultaneously.
• To overcome the computational demands and accuracy limitations of conventional grid-based approaches.

Main Methods:

• A Transformer deep learning architecture was developed and trained exclusively on simulated acoustic data.
• The model predicts spatially clustered source components, enabling individual source strength determination through integration.
• Strategies for reducing training effort across different frequencies were investigated.

Main Results:

• The method demonstrated fast and accurate source mapping for up to ten sound sources across various frequencies.
• Performance was validated against established methods like CLEAN-SC and sparse Bayesian learning using experimental data.
• The approach successfully characterized individual source strengths by integrating predicted cluster components.

Conclusions:

• The proposed deep learning method offers a computationally efficient and accurate alternative for grid-free sound source characterization.
• The single-model architecture effectively handles an unknown number of sources, outperforming traditional and some existing grid-free techniques.
• This approach holds significant potential for applications requiring precise acoustic source identification without grid-based constraints.