Skip to content
CrowLingo

FIG 2.1 — The Crow · Repertoire Atlas

The vocal map.

Explore 15 American crow recordings. Select a bright dot to listen, inspect its sound picture, and check the source. Choose a legend label for an introduction to that group. The nine groups come from source metadata; map positions and context weights are illustrations, not measurements from an AI model.

New to crow calls? Start with a short listen →Compare caw and rattle examples →
Loading 795 vocalizations…

Inline glossary

The Atlas, in plain English.

Embedding
A list of numbers a model uses to describe a sound. This Atlas has not computed embeddings.
Latent space
A way to compare those lists of numbers. In measured data, nearby points may be similar; this map’s positions are illustrative.
UMAP
A method that turns many measurements into a two-dimensional picture. It was not run on these recordings.
Cluster
A group of similar examples in an analysis. Here, labels are editorial groups based on source notes.
Bridge point
A point between groups. These bridge points are synthetic illustrations, not measured variation.
Context
What was happening around the bird. These files have no synchronized observation logs; the displayed weights do not establish its behaviour.

The deep methodology lives at Latent Space 101 and NatureLM-audio.

Frequently asked

What people ask about this.

What is the vocal map?

The vocal map is an educational illustration. Bright dots open archival recordings with real spectrograms. Groups are assigned from source metadata; all coordinates and context weights are illustrative. No embedding model, UMAP, or classifier was run on this corpus.

What is UMAP?

UMAP is uniform manifold approximation and projection — a non-linear dimensionality reducer that flattens high-dimensional embeddings to two dimensions while preserving local neighborhood structure. UMAP is described as a possible method; this atlas uses seeded illustrative coordinates instead.

Are the cluster boundaries discovered or assigned?

The current atlas uses nine editorial groups inspired by descriptive literature. Recording assignments use source metadata. These groups were not discovered by HDBSCAN and are not a validated count of the species' vocal categories.