Alba Márquez-Rodríguez

Bioacoustic foundation models are increasingly used in wildlife monitoring programs, yet their performance has rarely been evaluated at global scales. Here, we assessed the performance variation across species and regions of Perch, a foundation model capable of recognizing the sounds of 9,706 bird species worldwide. We then compared its performance against BirdNET, a popular foundation model for automated bird acoustic recognition that covers 6,421 bird species. We analyzed 4,224 expert-annotated 1-min recordings from 67 datasets across six continents, comprising 89,061 bird vocalizations from 1,102 species. Perch performance varied markedly among regions, performing best in Europe and North America, followed by Central and South America, and was substantially worse in Africa, Asia, and Oceania. These geographical patterns closely resembled the ones found for BirdNET, highlighting persistent regional biases in both models. Overall, Perch detected a comparable number of vocalizations and species to BirdNET, but showing higher precision. Perch achieved higher precision in 94% and 84% of datasets at the vocalization and dataset (species-inventory) levels, respectively. Perch performance was highly sensitive to confidence score thresholds, with increasing thresholds substantially reducing false positives while markedly decreasing recall. Nonetheless, the impact of threshold selection varied greatly among species, suggesting the need for species-specific assessments before applying Perch in ecological studies. Our results show that Perch is a powerful tool for automated bird monitoring, outperforming BirdNET in several datasets, primarily because of its higher precision. However, important geographical biases remain, particularly in Africa, Asia, and Oceania, highlighting the need for additional training data and targeted model development to improve automated acoustic surveys. Further research should also assess the influence of user-defined settings for improving Perch performance.

1. Passive acoustic monitoring (PAM) enables non-invasive sampling of wildlife across broad spatial, temporal and taxonomic scales. Its ongoing and widespread use has generated unprecedented volumes of acoustic data, shifting the primary bottleneck from data collection to the storage, processing, integration, and interpretation of PAM outputs. Although many software tools exist to address these challenges, differences in their design, scope, and usability often create fragmented and complex analytical workflows. To identify the key barriers and opportunities shaping the implementation of PAM surveys, we conducted a structured expert solicitation involving 30 international practitioners working across terrestrial and aquatic ecosystems. Experts identified and ranked their most critical pain points in current PAM workflows, spanning data storage, processing, and interpretation. The top challenge identified related to accurate species identification using deep learning and artificial intelligence (AI) models, especially in noisy soundscapes or for underrepresented taxa. Eight additional priority challenges included workflow fragmentation, limited availability of user-friendly analytical and visualisation tools, uneven access to software, manual validation bottlenecks, computational constraints, and difficulties in data handling, standardisation, and sharing. Participants also proposed practical mitigation strategies for these priority challenges, supported by step-by-step guidance to help overcome key barriers. Together, these insights provide a roadmap toward more scalable, open-access, and collaborative software systems, which are increasingly essential to realise the full potential of PAM in global biodiversity monitoring.

Passive acoustic monitoring (PAM) enables non-invasive sampling of wildlife across broad spatial, temporal and taxonomic scales. Its ongoing and widespread use has generated unprecedented volumes of acoustic data, shifting the primary bottleneck from data collection to the storage, processing, integration, and interpretation of PAM outputs. Although many software tools exist to address these challenges, differences in their design, scope, and usability often create fragmented and complex analytical workflows. To identify the key barriers and opportunities shaping the implementation of PAM surveys, we conducted a structured expert solicitation involving 30 international practitioners working across terrestrial and aquatic ecosystems. Experts identified and ranked their most critical pain points in current PAM workflows, spanning data storage, processing, and interpretation. The top challenge identified related to accurate species identification using deep learning and artificial intelligence (AI) models, especially in noisy soundscapes or for underrepresented taxa. Eight additional priority challenges included workflow fragmentation, limited availability of user-friendly analytical and visualisation tools, uneven access to software, manual validation bottlenecks, computational constraints, and difficulties in data handling, standardisation, and sharing. Participants also proposed practical mitigation strategies for these priority challenges, supported by step-by-step guidance to help overcome key barriers. Together, these insights provide a roadmap toward more scalable, open-access, and collaborative software systems, which are increasingly essential to realise the full potential of PAM in global biodiversity monitoring.

BirdNET is a popular machine learning tool for automated recognition of bird sounds. However, evidence on how to optimize its settings for accurate bird monitoring remains limited. Here, we evaluate how BirdNET settings influence model performance in identifying bird vocalizations and characterizing bird communities, using 4224 1‐min recordings from 67 recording locations worldwide. Giving equal importance to recall and precision, a low confidence score threshold (0.1–0.3) appears optimal for detecting bird vocalizations, whereas higher thresholds (around 0.5) are more suitable for characterizing bird communities. Based on our findings, we recommend increasing the Overlap parameter from its default value of 0 to 2 s, as this consistently improves BirdNET performance in detecting both bird vocalizations and species presence. The effect of the Sensitivity parameter varied across regions. However, a value of 0.5 maximizes global performance for community‐level analyses across all confidence thresholds, and a value of 1.5 generally yields better results for vocalization‐level studies, particularly at low confidence thresholds. Our findings offer practical guidance for selecting BirdNET settings in passive acoustic bird surveys, enhancing both the identification of bird vocalizations and the characterization of bird communities.