Identificación acústica de especies a partir de audio, utilizando algoritmos de aprendizaje automático

Abstract

Abstract: This project focuses on the field of machine learning, specifically on developing an automatic learning model for species recognition from audio recordings. The initiative originated from a group of conservationists belonging to the El Silencio Biological Reserve, who are responsible for biodiversity recovery and the protection of endangered species. In this context, various species are monitored as part of conservation programs. With the objective of automating biodiversity monitoring, this project employs a dataset obtained from the Kaggle platform comprising 28,564 total samples, containing short recordings of individual sounds from 206 species present in the Central Andes of Colombia. The dataset shows a distribution of birds with 27,648 samples, representing 96.79% of the dataset, followed by amphibians with 583 samples (2.04%), mammals with 178 samples (0.62%), and insects with 155 samples (0.54%). The study experiments with architectures based on Convolutional Neural Networks, specifically EfficientNet-B1, ResNet50, and PANNs, which enable the identification of acoustic patterns through a smartphone microphone for subsequent classification.