Beamforming informed independent low-rank matrix analysis for sound source enhancement in unmanned aerial vehicles

Abstract

This study proposes the integration of informed, supervised, and blind sound source enhancement approaches for unmanned aerial vehicle (UAV) applications. The proposed method incorporates a beamformer, representing the informed approach, a pre-recorded noise database for the supervised approach, and independent low-rank matrix analysis (ILRMA) for the blind approach. This method aims to improve sound source enhancement performance while addressing the permutation ambiguity problem of the output channels inherent to ILRMA. The method leverages the fixed spatial relationship between the UAV’s propellers and microphones to capture spatial information of the noise generated the propellers. This is achieved by deriving a noise covariance matrix from pre-recorded propeller noise signals and incorporating it into a general eigenvalue beamformer to effectively suppress these noises. The filter weights of the beamformer are then used to inform the spatially regularised ILRMA, guiding the algorithm with additional spatial information of the sound sources. Experimental results demonstrate significant performance improvements, including a 13 dB increase in the source-to-distortion ratio and a 0.21 point increase in the short-time objective intelligibility score. The proposed method outperforms both the original ILRMA and the beamformer in most scenarios and effectively addresses the global permutation ambiguity problem.