Bird sound detection with binarized neural networks

Abstract

By analysing the behavioural patterns of bird species in a specific region, researchers can predict future changes in the ecosystem. Many birds can be identified by their sounds, and autonomous recording units (ARUs) can capture real-time bird vocalisations. The recordings are analysed to see if there are any bird sounds. The sound of a bird can be used for further analysis, such as determining its species. Bird sound detection using Deep Neural Networks (DNNs) has been shown to outperform traditional methods. DNNs, however, necessitate a lot of storage and processing power. The use of Binarized Neural Networks (BNNs) is one of the most recent approaches to overcoming this limitation. In this paper, a bird sound detection architecture based on the XNOR-Net variant of BNN is used. Performance analysis of XNOR-Net in terms of the number of hidden layers used was performed, and the configuration with the highest accuracy was built. The system was tested using Xeno-Canto and UrbanSound8K datasets to represent bird and non-bird sounds, respectively. We achieved 96.06 per cent training accuracy and 94.08 per cent validation accuracy. We believe that BNNs are an effective method for detecting bird sounds.