What Is cosine.club and What Does It Do?

cosine.club is an electronic music similarity search engine. You give it a track, and it returns other tracks that sound close to it, matched by analyzing the audio itself rather than tags, genres, or listening history. It searches a catalog of over two million electronic tracks, and per the site's own description it is free and requires no account. It is most useful when you are trying to find more music in the same sonic vein as an obscure, untagged, or barely-played record — the exact case where metadata-driven recommenders tend to fail.

How the matching actually works

The mechanism matters, because it explains both what the tool is good at and where it breaks down.

1. Audio is converted to a vector. cosine.club uses the discogs-effnet model from Essentia, a neural network trained to recognize 400 Discogs styles from two-second slices of a spectrogram. The embedding it produces is a vector of 1,280 dimensions describing how a track sounds.
2. Every catalog track already has its vector stored. That's the "over two million" figure — the comparison set is precomputed.
3. Your track is run through the same model. When you paste a track, it goes through the same pipeline, and the result is compared against the whole catalog using cosine similarity (a standard measure of how close two vectors point in the same direction).

One consequence worth knowing: because the model was trained on Discogs styles, those 1,280 dimensions encode era, place, and scene as well as raw sound. So "similar" here means similar in the way a genre-aware listener would group records — not just similar timbre.

What it is not

It is not Shazam. Shazam fingerprints a recording, which is why it survives crowd noise and works from a phone mic in a club. cosine.club matches by sound character, so identification is a side effect at best — the site itself frames identification as "sometimes" possible, not the core function.

It is also not Spotify or Last.fm. Those systems build recommendations from metadata and co-listening behavior. That works well for music lots of people already play and badly for everything else: a record needs tags, plays, and co-listening data before it can be matched at all. cosine.club reads the audio instead, so a 300-pressed white label with no tags and almost no plays sits in the catalog on the same terms as anything else.

Who it is for

• Diggers chasing a specific sound. You have one track and want more like it, including records that share no tags with it.
• Listeners into obscure electronic music. The less metadata a track has, the bigger the advantage over tag-based recommenders.
• Anyone exploring by texture rather than genre label. If you care about how a record feels more than what shelf it sits on, the audio-first approach fits.

It is less useful if you want mainstream discovery, playlist-style radio, or track identification from a noisy recording.

What you can do on the site

Beyond single-track search, the page surfaces a few entry points:

The FAQ is organized as plain text files under /help/faq/ — HOW_IT_WORKS.txt, SEARCH.txt, TRACK_DATABASE.txt, PLAYLISTS.txt, TROUBLESHOOTING.txt, API.txt, and others — which is where the technical detail above comes from.

Practical notes before you start

• No account is needed to search, per the site's description. There is a log-in option on the page, which implies account features exist, but the core search is presented as free and account-free.
• The catalog skews electronic. This is not a general-purpose music search engine.
• Expect era and scene clustering. Because the embedding carries era and place, results for a 90s techno track will tend to stay in that neighborhood rather than jumping across decades — useful if that's what you want, limiting if it isn't.
• Treat "similar" as a starting point, not a verdict. Vector similarity is a ranking signal; your ears still decide.

If your problem is "I have this one record and I want more that sound like it, and the usual recommenders keep giving me nothing," cosine.club is built for exactly that. If your problem is "identify this song playing in the background," use a fingerprinting app instead.