Clean Data, Simple Models: A Practical Audio Preprocessing Approach for Multi-Species Sound Classification

Juan J. Diaz, Santiago J. Vargas, Isabella Bermón, Maria J. Guerrero

Keywords: Preprocessing, Embeddings, Bioacoustics, Passive Acoustic Monitoring, Sound Identification

TL;DR: Clean audio beats complexity, our preprocessing pipeline lets simple models rival deep baselines for BirdCLEF 2025 in noisy, resource-limited settings.

Abstract: We present a lightweight audio-preprocessing pipeline that boosts simple classifiers for multi-species sound identification in Colombian soundscapes. Developed for BirdCLEF 2025 and evaluated on recordings from Reserva Natural El Silencio (Magdalena Medio Valley), the pipeline isolates vocalizations, removes silence, and filters noise to produce cleaner BirdNET embeddings. We train MLP and CNN models on raw vs. cleaned inputs. Results in multi-taxon species classification show that improving signal quality can offset model complexity, where a cleaned-input MLP matches or surpasses deeper baselines with modest compute. This underscores the value of preprocessing for bioacoustic monitoring in noisy, resource-limited settings and demonstrates that robust baselines can be built with accessible computing resources common in biodiversity-rich developing countries.

Submission Number: 31