Semi-blind source separation for unmanned aerial vehicle audition

This paper presents a semi-blind source separation (BSS) method tailored for sound source enhancement for audio recording systems mounted on unmanned aerial vehicles (UAVs). This method capitalises on recordings of UAV ego-noise to supervise the independent low-rank matrix analysis (ILRMA) algorithm. Through the integration of spatial and noise source supervisors, ILRMA is transformed from a blind to a semi-blind method, substantially enhancing sound source separation performance in UAV settings. The spatial supervisor effectively addresses the global permutation problem in BSS within input signal-to-noise ratios (SNRs) ranges of 0 to -30 dB. Concurrently, the noise source supervisor leverages the UAV's dominant ego-noise to predetermine the BSS solution for noise components, leading to improved performance. Comprehensive tests using generated and recorded target signals demonstrate significant performance improvements, including an 18 dB increase in source-to-distortion ratio, a 20 dB increase in signal-to-noise ratio, a 0.22 score improvement in short-time objective intelligibility, and a 0.5 dB improvement in cepstral distance.