Made by Dario Baldini.
View the music clustering code on GitHub.
What you see is the 3D map of the vector representation of a music collection.
A pretrained Discogs-EffNet audio model analyzes the track and generate 2560-value embedding for the song. These features summarize aspects of the audio learned by the model; they are not manually assigned genre labels.
The embeddings are standardized and organised by their feature similarity using cosine distance. Dense groups are then found in the high-dimensional semantic space without requiring a preset number of clusters. Songs that do not fit a sufficiently dense group are marked as outliers. A song's cluster probability indicates its membership strength.
Cluster names are descriptive summaries, not genre classifications: they are defined by identifying the most distinctive artists in each cluster.
A separate projection reduces the song embeddings to three dimensions for display. Each point is a song, and nearby points generally have more similar audio embeddings in the high-dimensional semantic space. The projection is a visual aid: distances and directions in 3D are approximate and should not be interpreted as exact measurements.
Navigate the map
- Drag to rotate the 3D map, scroll to zoom, and hover over a point to see the song's title, artists, album, release date and cluster.
- Choose Cluster to color songs by their assigned cluster, Artist to color them by artist, or Date to color them by release year.
- In Artist mode, each artist has a distinct color. In Date mode, toggle whether the oldest and newest 2% of release dates are ignored when setting the color scale.
- Use the legend to show or hide groups. On phones and tablets, tap Legend to open the scrollable group list. The song count is displayed in the controls.
Find and explore songs
- Search by song title, artist, album or cluster in the panel on the left. Searches can combine terms; use Reset to clear the search and selected song, or Random song to clear the search and open a random song card.
- Select a search result or click a point to open its song card. The selected song is highlighted on the map. By clicking on a point, the YouTube player is also launched with that song.
- The song card shows the available song information. Click eligible artist or cluster values to search for exact matches, and follow available external links for more details.
- Use Play song to open the song's YouTube player, or Share song to share or copy a link that opens directly to that song.
- Explore the listed most similar and most different songs. Select any title in those lists to open its song card and continue exploring.