A corpus of bird sounds from Quindio and its application for passive acoustic monitoring through neural networks

Abstract

The biodiversity of a region is an invaluable asset that requires continuous efforts for its preservation and study. Particularly, for the region of Quindio in Colombia it is highlighted its richness in native birdlife, which enriches its natural landscape, and significantly contributes to the country biological diversity. The observation and study of those species becomes an essential task for the conservation and knowledge of the region. In this context, this paper proposes a corpus of sounds with a ML processing support. Datasets that were used include audios of 170 bird species, which corresponds to approximately 30% of the bird species identified in Quindio. Audios of human voices, silences and noises also were included. For signal processing, the sliding window feature extraction technique is used to analyze and classify bird sounds. Additionally, three neural networks were trained to evaluate the corpus, the first being a convolutional network. From the results of this network , two additional networks were trained, one of which was another convolutional network, while the second was based on the transformers architecture. These networks were trained with the categories that showed performance with an F1-score metric equal to or greater than 0.30 in the first convolutional network. The results obtained show precision levels of 0.55, 0.53 and 0.65 respectively. Network based on Transformers demonstrated better performance in classifyingthe sounds of native birds of Quindio. A proof of concept was carried out on this network with audios of the species Saffron Finch (Sicalis flaveola), reaching an accuracy of 65.74%. These results offer a baseline for future research in the field of bird sound classification, thus promoting the conservation of regional avifauna.