Audio Segmentation to Build Bird Training Datasets

Abstract

To create a bird classification model, it is necessary to have training datasets with thousands of samples. Automating this task is possible, but the first step is being able to segment soundscapes by identifying bird vocalizations. In this study, we address this issue by testing four methods for audio segmentation, the Librosa Library, Few-Shot Learning technique: the BirdNET Framework, and a Bird Classification Model called Perch. The results show that the best method for the purpose of this work was BirdNET, achieving the highest values for precision, accuracy, and F1-score.