Deep learning based bird sound classification using customized convolutional neural networks

The classification of sounds produced by birds is very helpful in ecological monitoring, research on biodiversity, and conservation activities. It can be quite difficult to recognize bird species by their calls, especially in places with many overlapping sounds or noise. This paper presents an approach for automatically classifying bird sounds with deep learning using customized Convolutional Neural Networks (CNNs). The method extracts MFCCs to capture important discriminative features from the Xeno-Canto dataset, which consists of more than 3000 audio files from 114 bird species. Three Conv1D blocks with batch normalization, max-pooling, and dense layers process these features, producing a SoftMax output of 114 units. The data is pre-processed with optimized pipelines in TensorFlow, including caching, shuffling, and prefetching, which improves model performance as well as the time taken to train the model. With the proposed approach, the new model based on CNN reached an accuracy of 93.4%, which is a significant improvement than that achieved from prior methods. The proposed model is capable of automatically and accurately identifying birds by their songs, which can greatly benefit eco-biology studies, conservation, and monitoring research.