Can AI name an East African bird from its song?

We searched iNaturalist for bird recordings from South Sudan, Djibouti, Eritrea, Somalia, Tanzania, Rwanda, Burundi, Kenya, Uganda and Ethiopia, and gathered 1,281. This report uses the 589 that iNaturalist marks Research Grade, meaning the community has agreed on the species. The rest are still waiting for that agreement and are left out.

Where the 589 come from: Kenya 263, Uganda 121, Tanzania 116, Rwanda 36, Ethiopia 17, Somalia 3, South Sudan 1, Burundi 1. 31 have no country on the record.

A model’s first answer is the bird it is most sure of anywhere in the recording. It counts as a match when that is the species the community named. A model that gave no answer counts as a miss, and a recording named down to a subspecies is matched by its species.

The sound and photograph counts come from eBird’s Macaulay Library, which we copy every week for the same countries. They are as of 27 September 2026, cover East Africa only, and count a bird’s named sub-groups with the bird. Recordings in that archive that are not of a bird species, such as chimpanzees and frogs, are left out.

A model may have heard some of these recordings before. Perch v2 was trained partly on Research Grade sound from iNaturalist, downloaded in March 2025, and 422 of the 589 recordings were made before then. We did not check which ones it was trained on, or what BirdNET was trained on, so a match here can be a recording the model already knew.

The recordings are the ones that fit our rules, not a random sample, and a few keen recordists supplied many of them. So these results describe this collection. They cannot crown a winner for every place and every bird.