Automated Bird Species Identification using Audio Signal Processing and Neural Network

Many bird species are rare nowadays, and when they are found, they are difficult to classify. As an example, in various scenarios, birds include different sizes, forms, colors, and a person's viewpoint from different angles. Although domain specialists can classify birds manually, with increasing volumes of data, this becomes a tiresome and time-consuming procedure. Using our approach, we can reliably and quickly identify bird species. It is now feasible to track the number of birds as well as their activity using automated bird species recognition and machine learning algorithms. Convolutional neural networks (CNN) were chosen above standard classifiers such as SVM, Random Forest, and SMACPY. For this system, we used the “BirdCLEF 2021” dataset from Kaggle. The input dataset will be preprocessed, which will involve framing, silence removal, and reconstruction, which will be supplied as input to a convolutional neural network, followed by CNN modification, testing, and classification. To avoid overfitting, we add a dropout layer. Preprocessing includes importing the Librosa library. MFCC is a program that extracts distinct characteristics from audio files (Mel-Frequency-Cepstral-Coefficients). The MFCC summarizes the frequency distribution over the window size, allowing for sound frequency and temporal analysis. The result is then compared with respect to the pre-trained data, and output is shown, and birds are classified based on their classes. 

Keywords: Bird sound classification, Classification, Convolutional neural networks (CNN), Dataset, Librosa, Mel-frequency-cepstral-coefficients (MFCC), Pre-processing, Random forest, SMACPY.