Speaker
Description
Acoustic monitoring has become a widely used approach to study ecosystems. However, the large amount of acquired data requires fast processing techniques to efficiently analyze soundscapes. For that, sound detection and classification algorithms, such as BirdNET, have been developed. However, the performance of BirdNET, quantified by the confidence score, is closely linked to the Signal-to-Noise Ratio (SNR) of the audio recordings. Besides, the use of Spherical Microphone Arrays (SMA), which enable rotationally invariant Spherical Harmonics (SH) beamforming, allows for directional filtering of the soundscape and sound source localization in any direction. In this work, natural soundscape analysis is performed using a SMA for recordings. In particular, the night call of the little owl (Athene noctua) was automatically searched for. Firstly, sector-based filtering is carried out to provide a set of spatially filtered signals. These signals are then analyzed using BirdNET to provide a list of sound events. Secondly, for each sound event, the CLEAN-SC deconvolution beamforming algorithm is applied to the whole 3D sound scene captured by the SMA. Directional signals are finally extracted in the direction of the sound sources and analyzed through BirdNET. An improvement of over 200% in the number of little owl detections in a natural environment has been achieved by means of sector-based filtering in comparison to a monophonic signal. After extracting directional signals, the confidence score of BirdNET has been raised by more than 50%. These results suggest that the proposed method improves the robustness of BirdNET.