Sound source localization and speech enhancement with sparse Bayesian learning beamforming

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Conventional beamforming methods for speech localization and enhancement have limited resolution, particularly with fewer microphones.
• Estimating direction-of-arrival (DOA) is crucial for reconstructing sound sources from noisy microphone array recordings.

Purpose of the Study:

• To address the low-resolution limitations of conventional beamforming for speech localization and enhancement.
• To apply sparse signal reconstruction techniques, specifically sparse Bayesian learning (SBL), to improve DOA estimation.

Main Methods:

• Formulating DOA estimation as a sparse signal reconstruction problem.
• Utilizing sparse Bayesian learning (SBL) with a hierarchical two-level Bayesian inference framework.
• Adaptively learning hyperparameters to auto-regularize the inference towards sparse, robust estimates.

Main Results:

• SBL beamforming achieves high-resolution DOA maps, outperforming traditional methods.
• Demonstrated superior performance with correlated or non-stationary signals.
• Achieved effective speech separation and enhancement for speech signals.

Conclusions:

• Sparse Bayesian learning provides a robust and high-resolution approach for DOA estimation in speech processing.
• SBL enables significant improvements in both speech enhancement and speech separation capabilities.
• The adaptive hyperparameter tuning in SBL enhances the reliability of sparse signal reconstruction.