Bird-sound recognition applications can appear highly convincing, but how reliable are their results in real field conditions? This report compares four bioacoustic recognition models in a Finnish rural soundscape: Muuttolintujen kevät v4, BirdNET 2.4, BirdNET 3.0 and Google DeepMind’s Perch 2.0.
The comparison revealed substantial species-specific differences. Muuttolintujen kevät v4, trained for Finnish conditions, produced the most consistent overall results and detected species such as Hooded Crow, Canada Goose, swallows and swifts better than the other models. The new BirdNET 3.0 showed strong potential as a global model, while Perch 2.0 stood out through its ability to recognise other animals and environmental sounds in addition to birds.
The results also exposed important pitfalls in automated identification. A high confidence score—or even agreement between several models—did not always mean that a detection was correct. In this summer dataset, for example, every checked Long-eared Owl and Eurasian Eagle-Owl detection was false. On the other hand, a single detection during migration may still represent a genuine fly-over bird.
The report explains what the models identified correctly, where they failed, and how bioacoustic results should be filtered and verified in practical monitoring. It includes model-specific strengths and limitations, species-group comparisons, detailed data tables and a recommended workflow for automated nature
Download the full report to explore the results.
