A Bilinear Source Separation, Dereverberation, and Background Noise Suppression Algorithm for Augmented Reality Applications

This paper introduces a new bilinear algorithm that combines distortionless response beamforming with weighted prediction error using the Kronecker product operator to improve speech signal quality in different acoustic environments. The algorithm utilizes recursive least squares cost functions and handles real dynamic, noisy, and reverberant conditions effectively. Compared to a recently proposed approach, our approach outperforms separating the desired source from undesired sources while providing dereverberation and noise suppression. It is also preferable regarding the quality of the enhanced signals it produces. In contrast, it exhibits a more considerable desired signal distortion and reduced intelligibility. The algorithm’s computational efficiency and robustness make it suitable for real-time applications, as validated using real recordings from the SPeech Enhancement for Augmented Reality (SPEAR) challenge.