Deezer uses LLMs to caption millions of playlists daily

Deezer has published a paper detailing its production system for automatically captioning millions of personalized playlists using large language models (LLMs). Deployed in 2025, the system generates concise, natural-language descriptions for each Daily Mix playlist—the feature that serves millions of users daily. By taking diverse data sources (e.g., track metadata, user listening history) and feeding them into an LLM with controlled generation parameters, Deezer creates captions that explain what a playlist contains and why it might be relevant to the listener.

The impact on user engagement has been significant: the paper reports that adding these semantic captions to an otherwise unchanged recommendation algorithm changed how users perceive and interact with suggested playlists. This work, accepted at ECML-PKDD 2026, demonstrates a practical application of LLMs to improve the user experience in music streaming at scale, without altering the underlying recommendation engine.