BAT-CNN: BirdNet Assisted Training for CNN

Studies on bird sound classification have become increasingly popular in recent years. Birds serve as valuable ecological indicators, and the threat to bird species populations is a serious concern. The emergence of artificial intelligence has been a milestone in species classification, making the classification easier and more accurate. For the acoustic classification of bird species, a number of works have been reported in the literature but there are still challenges. In this, we propose a deep learning approach for bird sound classification. A dataset consisting of recordings of 10 bird species is used for classification. The dataset is sourced from recordings on the Xenocanto website. Many bird audio recordings are available for studying their existence, but classifying these recordings is challenging due to the overlapping calls from multiple birds. A different approach is used in the proposed work for preprocessing the downloaded audio. BirdNet, which can recognize over 6500 birds, is employed for separating the audio of the corresponding species from the overlappings of the multiple species present in the downloaded audio files. The BirdNet separated recordings of the 10 species are then converted into spectrograms. The obtained spectrograms are then fed into a CNN model for classification. The model achieved an average accuracy of 90% for the ten bird species.