BeamLearning: An end-to-end deep learning approach for the angular localization of sound sources using raw...

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Multichannel signal processing for sound source localization is a long-standing research area.
• Deep learning has advanced machine hearing, prompting its application in source localization.
• Traditional methods often rely on simplifying assumptions about acoustic environments.

Purpose of the Study:

• To introduce BeamLearning, a multiresolution deep learning approach for sound source localization.
• To process raw time-domain acoustic signals from microphone arrays, avoiding model-based simplifications.
• To evaluate BeamLearning's effectiveness in real-time 2D sound source localization within challenging environments.

Main Methods:

• Developed a multiresolution deep learning architecture (BeamLearning) for raw acoustic signal processing.
• Utilized graphics processing unit (GPU)-based computation of room impulse responses with fractional delays for dataset generation.
• Tested the BeamLearning network using both synthetic and experimental datasets.

Main Results:

• BeamLearning demonstrated superior performance in 2D sound source localization compared to wideband MUSIC and SRP-PHAT methods.
• The approach showed significant improvements in localization accuracy under heavy measurement noise and reverberation.
• BeamLearning exhibited enhanced computational efficiency.

Conclusions:

• The BeamLearning deep learning approach effectively localizes sound sources using unprocessed acoustic data.
• It offers significant advantages over traditional methods in noisy and reverberant conditions.
• The method shows promise for real-time acoustic signal processing applications.