Bird Species Recognition in Soundscapes with Self-supervised Pre-training

Biodiversity monitoring related to bird species is often performed by identifying bird species in soundscapes recorded by microphones placed in the birds’ natural habitats. This typically produces a large amount of unlabeled data. While self-supervised machine learning methods have recently been successfully applied to computer vision and natural language processing tasks, state-of-the-art automatic approaches for bird species recognition in audio recordings mainly rely on transfer learning using pre-trained ImageNet models. In this paper, self-supervised learning is leveraged to improve bird species recognition in soundscapes. Specifically, we use a novel self-supervised approach to pre-train a self-attention neural network architecture on the target domain to take advantage of the vast amount of unlabeled and weakly labeled data. Experiments on data sets from different recording environments show the effectiveness of our approach. In particular, self-supervised pre-training on the target domain improves the cross-domain recognition quality.