Deep Learning-based Automatic Bird Species Identification from Isolated Recordings

Birds play an extremely important role in an ecosystem, identifying bird species in audio recordings is challenging and has high research value. This paper aims to develop an effective bird call classification for isolated recordings (single-label) approach using various deep learning architectures, namely convolutional neural networks (CNN), deep neural networks (DNN), and transfer learning schemes. Transfer learning models have been widely used in a variety of deep learning applications. The performance of transfer learning models such as ResNet50, VGG-16, and InceptionResNetV2 has been compared to the acoustic MFCC-DNN methodology. On the Xeno-canto (XC) online bird audio dataset, the presented methods are tested. The dataset comprises ten species with 1078 audio tracks. The classification accuracies of 96.3%, 93.7%, and 91.9% are reported for ResNet50, CNN, and VGG-16, respectively, and outperform with the acoustic signal-based MFCC-DNN methodology.