Why people need to listen to recordings when AI has already identified the species, and how to read automatic detections.
How it works
- Recordings are processed by classifiers: BirdNET 2.4, Perch v2 and Chirpity models for nocturnal migration. Each detection has a species and a model confidence score.
- The system selects fragments for checking.
- A person listens to the fragment, looks at the spectrogram and answers "correct" or "incorrect".
- From these answers the platform calculates how accurate the model is for each species and from which confidence threshold it can be trusted.
Why people are needed
AI accuracy by sound varies strongly between species. As of 19 October 2026 it has been checked for 46 species, and only 9 of them reach 90 % or more. Eastern tree frog, marsh frog, corncrake or great reed warbler are recognised almost perfectly, while common spadefoot or pool frog almost never. A species can be among the most frequent by number of detections while nearly all those detections are false.
How to read automatic data
- Before drawing a conclusion about a species, check whether the model's accuracy for it is known. The conclusion must be about the animal, not about the behaviour of the classifier.
- Compare locations, seasons and years only at the same confidence threshold. A low threshold adds false "rare" species, a high one loses quiet and distant calls.
- Compare only data with the same recording schedule. See PAM field methods.
- Use a week as the unit for detection estimates. A bird sings from the same branch every morning, so neighbouring days are strongly dependent.