Promise and pitfalls: Variable performance of AI-assisted passive acoustic monitoring of birds and frogs using BirdNET and VicFrogNET

This study introduces VicFrogNET, the first open-source multi-species classifier for Australian frogs, built using transfer learning from BirdNET on a small, curated dataset of 16 native species. The authors assess the frog detection performance of VicFrogNET, while separately applying the existing BirdNET classifier to assess bird detection performance. Using more than 1.5 TB of passive acoustic recordings from 60 freshwater wetlands, the study analysed 820,740 bird detections and 2,419 frog detections, with more than 17,000 detections manually validated. Both classifiers showed strongly bimodal performance, with species generally detected either with very high or very low precision. VicFrogNET achieved high precision for most frog species despite limited training data, suggesting that curated, high-quality training data may be more important than quantity for targeted regional classifiers. BirdNET performance was more variable across Australian birds, requiring cautious interpretation. The study highlights the value of open, transparent benchmarking while emphasising the need for species-specific validation.