Deep learning for detection of bird vocalisations

Abstract

This work focuses on reliable detection of bird sound emissions as recorded in the open field. Acoustic detection of avian sounds can be used for the automatized monitoring of multiple bird taxa and querying in long-term recordings for species of interest for researchers, conservation practitioners, and decision makers. Recordings in the wild can be very noisy due to the exposure of the microphones to a large number of audio sources originating from all distances and directions, the number and identity of which cannot be known a-priori. The co-existence of the target vocalizations with abiotic interferences in an unconstrained environment is inefficiently treated by current approaches of audio signal enhancement. A technique that would spot only bird vocalization while ignoring other audio sources is of prime importance. These difficulties are tackled in this work, presenting a deep autoencoder that maps the audio spectrogram of bird vocalizations to its corresponding binary mask that encircles the spectral blobs of vocalizations while suppressing other audio sources. The procedure requires minimum human attendance, it is very fast during execution, thus suitable to scan massive volumes of data, in order to analyze them, evaluate insights and hypotheses, identify patterns of bird activity that, hopefully, finally lead to design policies on biodiversity issues.