Recuperação de paisagens sonoras: uma análise de técnicas de fusão de vetores de embeddings pré-treinados

Abstract

The retrieval of similar soundscapes is essential for bioacoustic and ecoacoustic monitoring, tasks that remain challenging due to the large volume of unlabeled data, environmental noise, and the complexity of acoustic scenes. This study proposes an efficient system that integrates embeddings extracted from a pre-trained deep learning model, combined with a noise reduction technique and feature vector fusion strategies to enable similarity-based retrieval in a vector database. We evaluated the system using bird, amphibian, and mammal recordings across four experimental methodologies, including a case study focused on endangered species. The results show that embedding vectors consistently outperform traditional MFCC features in capturing acoustic similarity and that the approximate search algorithm (HNSW) significantly improves both retrieval precision and query efficiency. Additionally, the system effectively retrieves recordings of the critically endangered species Crax alberti, allowing for the mapping of its geographic distribution and highlighting its potential for conservation planning.