BirdNET and Perch perform poorly for species inventories in Amazonia but can provide a proxy for community turnover

Global machine-learning classifiers for bird sounds are increasingly being integrated into acoustic biodiversity monitoring workflows. However, performance evaluations are concentrated in temperate regions, and are much less well understood in tropical systems. This gap is especially acute in hyperdiverse tropical forests, where test datasets are difficult to obtain. We assessed BirdNET v2.3 and Perch v2.0 predictions against 14,705 expert-labelled vocalisations from the Brazilian Amazon. Both classifiers showed weak classification performance, with precision–recall AUC of 0.07 and 0.12 for BirdNET and Perch respectively, driven primarily by extremely low recall, with mean recall scores of just 0.08 ± 0.09 (BirdNET) and 0.11 ± 0.13 (Perch) at the prediction scale. Performance improved when detections were aggregated across recording sites, with 64% of commonly detected species meeting minimum performance thresholds for at least one practical monitoring application. However, classification performance was significantly lower for passerines than non-passerines, with the most species-rich Amazonian families performing poorly, indicating a potential source of taxonomic bias. Classification performance significantly declined with increasing vocalisation frequency, but increased with signal-to-noise ratio and the amount of training data available in Xeno-Canto. Despite low species-level accuracy, classifier outputs preserved meaningful patterns of community composition, and predicted species richness tracked manually labelled richness at low confidence thresholds. Classifier outputs could therefore be treated as data-derived acoustic indices, rather than reliable species-level detections. Overall, our results suggest that global classifiers cannot yet support reliable species-level ecological inference in Neotropical rainforests, meaning studies using these approaches should complement them with extensive manual validation.