A deep learning method for grid-free localization and quantification of sound sources.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Accurate characterization of single point sources from microphone array data is crucial for various applications.
• Conventional beamforming methods have limitations in resolution and accuracy for source localization.

Purpose of the Study:

• To investigate the efficacy of deep neural networks (DNNs) for accurate single point source characterization using microphone array data.
• To develop a DNN-based method for estimating source coordinates and strength.

Main Methods:

• Utilized a residual network architecture, a deep learning model proven in image recognition.
• Applied the network to conventional beamforming maps to estimate source parameters.
• Trained and evaluated the model on microphone array data.

Main Results:

• The proposed DNN method accurately estimates both the position and strength of unknown single point sources.
• The method demonstrates fast processing speeds for source characterization.
• Position estimation accuracy achieved by the DNN method exceeds the grid resolution of traditional beamforming maps.

Conclusions:

• Deep neural networks offer a powerful and accurate approach for single point source localization and strength estimation.
• The residual network architecture is effectively transferable from image recognition to acoustic source characterization.
• This DNN-based technique provides a significant advancement over conventional methods in terms of speed and accuracy.