cosine.club

cosine.club utilizes a deep learning model, trained via contrastive learning on a large corpus of audio, to generate high-dimensional vector embeddings of songs. These embeddings encapsulate a rich representation of audio characteristics, effectively mapping complex acoustic features into a dense numerical space.

When a query track is input, its embedding is computed and compared against a database of pre-computed track embeddings using cosine similarity as the distance metric. This process identifies tracks with the most similar acoustic properties in the embedding space. cosine.club can identify musically similar tracks across genres and styles based on intrinsic audio features. This method captures tonal, rhythmic, and structural affinities between tracks, regardless of conventional categorizations or traditional metadata based approaches.

cosine.club was originally created to discover and find tracks that aren't available digitally or on platforms like Spotify. There are thousands of hidden gems that are ripped and uploaded to YouTube but aren't available anywhere else online. The scope of the database includes genres like House, Techno, UK Garage, Funk & Soul, etc., and not much mainstream or popular music (although some may slip through!). The database is sourced from Discogs.com and will be continuously updated as more releases are added to their database.

For tracks not found in the database, you can use the upload functionality by providing a YouTube or Bandcamp link. The system will analyze it and return similar tracks, enabling exploration of external tracks not in the system.