cosine.club VS JavaScript Tutorial

cosine.club's answer

Similarity is computed from the audio itself, not from tags, genres, charts or listening history. Every track is turned into a vector embedding by the discogs-effnet model from Essentia, and matches are found by cosine similarity between those vectors. Nothing in that process reads metadata, so results cross genre lines and turn up records that tag-based tools never connect.

The catalogue is the other half of it. Over two million tracks from Discogs, Bandcamp and SoundCloud, weighted heavily toward electronic and underground music rather than the charting catalogue most recommendation tools are built on.

If a track is not already indexed, paste a link to it on YouTube, Bandcamp or SoundCloud and the audio is analysed on the spot.

cosine.club's answer

Most similar-music tools work from metadata. Chosic and Music-Map build on tags and co-listening data, so a track has to be popular enough to have generated that data before it can be matched, and results tend to stay inside the genre they started in. Every Noise at Once maps genres rather than individual tracks, and has not been updated since 2023.

cosine.club reads the audio, so an obscure white label with no tags, no reviews and barely any plays sits in the index on the same terms as anything else. That is the case diggers actually run into, and it is the one metadata cannot handle.

It is free, search and similarity need no account, and there is a free public API.

cosine.club's answer

DJs, record collectors and crate diggers, mostly in electronic and underground music. House, techno, UK garage, jungle, breakbeat, electro, ambient, dub and disco.

The common thread is that they already know what they like. They are not looking for a personalised feed built from listening history, they want the neighbours of one specific record, and usually an obscure one.

A smaller group uses it to place unidentified rips. The nearest neighbours of an unknown track are often its label mates or its contemporaries, which is enough to work backwards from.

cosine.club's answer

It comes out of a problem that turns up constantly when you are digging for records. You hear a track you like and you want to find more that sounds like it.

Every tool that answered that question answered it with tags, charts or what other people were listening to. None of that works on the records worth digging for, because an obscure record does not carry enough of that data to be matched by it.

Audio embeddings get around the problem, since they only need the audio itself. That is what cosine.club is built on, over a catalogue of more than two million tracks.

cosine.club's answer

Go on the backend, with templ for HTML templating and HTMX on the frontend. Pages are rendered on the server and swapped in fragments, so there is very little client-side JavaScript and no SPA framework.

Similarity search runs on a vector database, text search on a separate search index, and PostgreSQL is the primary database.

Embeddings come from discogs-effnet, an Essentia model trained on the Discogs dataset.