The Zvuk music service has launched Atlas, an interactive map designed to help users explore their musical preferences. The system utilizes a two-dimensional model that categorizes tracks based on micro-genre axes and emotional mood axes, covering over 100 genres and nearly 50 distinct moods.
The technical architecture of Atlas integrates audio analysis, expert editorial labeling, and large language models. The EffNet neural network analyzes audio signals to determine probabilities across 400 micro-genres, which are then consolidated into 28 genre families. Editorial oversight is used to verify or refine these algorithmic results.
The mood axis is developed using LLMs based on the Russell model of musical emotions. This framework decomposes emotions into valence and energy quadrants, further refined by parameters such as sound texture, temperature, and chronotope, resulting in 12 emotional gradations per quadrant.
Personalized recommendations are generated by combining genre data with behavioral patterns, including listening depth, repeat plays, and collection additions. To access the Atlas feature, users must have at least 20 tracks in their collection, which triggers the automatic generation of their personal musical landscape.
Looking ahead, Zvuk plans to involve users as assessors to improve data labeling and intends to evolve the map into a continuous playback mode. According to CEO Artem Lebedev, the project aims to provide users with a visual representation of their musical world to facilitate discovery.
