The music service "Zvuk" has launched "Atlas" — an interactive map that shows a user's musical taste and helps them discover new compositions. To build it, the system analyzes not only genres but also the emotional characteristics of tracks.
At the core of "Atlas" is a two-dimensional model. One axis is responsible for micro-genres, the other for the mood of the music. Favorite compositions form "islands of sympathy" on the map, and the algorithm places personalized recommendations nearby.
Genres are determined by the EffNet neural network. It analyzes the audio signal and calculates the probability of a track belonging to one of 400 micro-genres. Then the system combines them into 28 large musical families.
Zvuk determines mood using language models. The algorithm considers valence and energy level, as well as sound characteristics, temperature, and chronotope. This results in 48 gradations of mood.
When selecting recommendations, the system also takes into account listener behavior: repeated listens, listening depth, and adding songs to the media library.
To get their own "Atlas," users need to have at least 20 compositions in their media library. After that, the service forms an individual musical landscape where new tracks appear next to familiar directions.