Source Enhancement with Different MVDR Beamformer Designs for Unmanned Aerial Vehicle Audition

This paper presents a sound source enhancement framework for unmanned aerial

vehicle (UAV) audition, designed to enhance sound sources in audio recording systems

mounted on UAV while reducing the recorded noise due to the UAV’s propellers. The framework

improves the robustness and performance of an existing beamforming-based sound

source framework in UAV audition, reducing its reliance on noise source information and

the direction of arrival (DOA) information of the target signal. Utilising the stationary spatial

relationship between the source position of UAV ego-noise and the onboard microphone

array, the framework estimates the noise covariance matrix using pre-recorded noise data.

Additionally, the framework incorporates a beamformer design that reassembles the characteristics

of a minimum variance distortionless response (MVDR) beamformer. It comprises

a general eigenvalue (GEV) beamformer with a compensation factor to partially emulate the

distortionless response characteristic of the MVDR beamformer. The beamformer is tailored

to be robust across a range of target positions, addressing the traditional limitations of an

MVDR beamformer and thereby improving its robustness in real-world applications. Additionally,

the framework adapts a post-processing technique, originally developed for singlerank

beamformers such as MVDR, to be compatible with general-rank beamformers such

as GEV, thus improving the overall sound source enhancement performance. Comprehensive

experimental evaluation demonstrates significant performance improvements, including

an 18 dB improvement in source-to-distortion (SDR) ratio and a 0.21 points improvement in

short-time objective intelligibility (STOI) score.