Listening to the planet

Someone stood in the Amazon as titi monkeys yelled across the canopy, in a Cairo market on a Ramadan night, beside a street vendor in Palestine calling out six pairs of socks for ten, and pressed record. Each point here is one of those moments. Two deep learning models have listened to each recording, and the sounds they hear as alike sit close together.

Which AI model is listening
Colour points by
Show the recordings on

What am I looking at?

When a deep learning model (AI) processes a recording, it turns the sound into a long list of numbers called an embedding. Recordings that sound alike end up with similar numbers.

Each list has hundreds of numbers, far too many to picture, so a mathematical transform called UMAP squeezes them into three dimensions while keeping similar sounds close together. Each point is one recording, placed by its embedding.

The two models are the interesting part. NatureBEATs was trained with a strong focus on animal sounds, BEATs on everyday audio, and switching between them rearranges the whole embeddings map. Seeing which sounds stay together and which drift apart is a nice, intuitive way into what embeddings actually capture.

Embeddings matter beyond this map. Among other things, they let scientists build detectors for new species from only a handful of examples, and our research* found that embeddings learned from birdsong transfer remarkably well to other animal sounds.

* Ghani, Burooj; Denton, Tom; Kahl, Stefan; and Klinck, Holger. (2023). Global birdsong embeddings enable superior transfer learning for bioacoustic classification. Scientific Reports 13, 22876.

Recordings come from Freesound, each credited to its recordist. In the backdrop is Dal Lake in Kashmir, photographed by Suhail Skindar Sofi (Wikimedia Commons, CC BY-SA 4.0).

© 2026 Burooj Ghani