Automated Bird Species Classification Using MFCC Features and Deep Neural Networks on the iBC53 Dataset

Birds are key players in ecosystem balance, and proper identification of species is necessary for biodiversity assessment and protection. In this paper we would like to introduce a deep learning based solution towards automatic classification of 53 Indian birds based on their vocalization using the iBC53 dataset. Its methodology uses Mel-Frequency Cepstral Coefficients (MFCCs) to represent informative bird collective calls and songs, and preserve the most significant frequency bands of each species. Multiple audio data augmentation techniques e.g. pitch shifting, time stretching and background noise addition during preprocessing were used to increase model generalization and decrease overfitting. Dense neural network architecture that borrows all the best properties from modern deep learning, including batch normalization and L2-regularization. The bird call classification accuracy of the proposed model was high upto95%. The system has a great potential for ecological field studies, in real-time species monitoring and conservation driven project solutions.