Explore the corpus
Every recording the commons holds, projected from 512 dimensions to two. Points that sit together sounded alike to the model. Click one, or paste a MusicBrainz recording or a hash.
Latest contributions
What you are looking at
512 numbers per recording, reduced to two by UMAP so they fit on a screen. The reduction is lossy and not unique: distances between distant clusters mean much less than the fact that a cluster is a cluster, and a different seed would rotate the whole picture. What survives is the neighbourhood structure.
Why names come from your browser
The corpus is keyed on the SHA256 of an AcoustID fingerprint and holds no titles. A contributor may claim a MusicBrainz recording id for a row; this page turns that id into a name by asking MusicBrainz itself, one request a second, from your machine — the server never resolves one.
Reproduce it
The projection is a committed artifact with a date on it, not something the server
computes — UMAP pulls in numba and llvmlite, and the instance is sized for a vector index.
packages/server/scripts/build_map.py regenerates it with a pinned seed.