Yves Raimond

Yves Raimond is Spotify’s senior vice president and general manager for personalization. His career spans music-data research, structured publishing at the BBC, and recommendation systems at Netflix, Google, and Spotify. His work at Spotify centers on generative personalization: giving listeners ways to explain, correct, and redirect what an algorithm thinks they want.

From connected music data to global recommendations

Raimond completed a PhD at Queen Mary’s Centre for Digital Music, where his early work concerned representing music and connecting information across databases. He helped develop the Music Ontology, a shared vocabulary for describing artists, recordings, performances, and their relationships, and created DBTune to publish interconnected music datasets. These projects made musical knowledge available as structured information that applications could reuse across catalogs.

At the BBC, he applied this approach to programme information. His work on BBC Programmes involved publishing structured data about broadcasts, alongside participation in the organization’s wider Semantic Web work. The challenge had expanded from describing music to connecting a broadcaster’s content on the web.

At Netflix, Raimond worked on personalization and held algorithm-engineering leadership roles. With Justin Basilico, he co-authored work on global recommendations during Netflix’s 2016 international expansion. Their approach combined signals across countries while accounting for local tastes, language, and differences in catalog availability. A viewer’s apparent lack of interest in a film could simply mean it had never been available to them—a distinction that matters when learning preferences from viewing behavior.

His collaborative research also examined the practical limits of deep learning in recommendation. In a Netflix case study, Raimond and five colleagues described how neural models initially struggled to improve on carefully tuned simpler systems. Richer, more varied input features helped make them useful, but better offline scores still needed to translate into improvements for real viewers. He subsequently worked on personalization for Google Search before joining Spotify, as he recounts in “Teaching LLMs to Speak Spotify,” presented with Jacqueline Wood.

Giving listeners a say in discovery

At Spotify, Raimond advocates personalization that responds to explicit intent alongside listening history. He describes a progression from manually curated playlists to ranked recommendations and then to experiences listeners can shape through natural language. A system that can interpret requests and explain its choices gives listeners more ways to influence discovery than a ranked list alone.

• Steerable recommendations: Spotify’s DJ and prompted playlists let listeners redirect a session or describe what they want to hear. The shared Large Taste Model combines prediction with reasoning about listening history and context, supporting recommendations and experiences that users can shape in real time. At the time of his 2026 presentation, Raimond reported that roughly one in four US Premium subscribers interacted with the system daily.

• An editable taste profile: Listening history contains ambiguity. Music played on a shared device may reflect someone else’s preferences, while a new interest may have no history at all. Raimond emphasizes letting listeners inspect and correct inferred preferences, as well as express aspirations such as exploring an unfamiliar genre or learning a language. These edits give the system information that behavioral prediction alone can miss.

• Catalog-grounded language models: Raimond co-authored Spotify’s NEO research, which combines natural language with semantic identifiers representing catalog items. The model can generate recommendations and explanations together, with item references tied to the catalog. He also co-authored GLIDE, a deployed podcast recommender that combines recent listening, contextual information, and longer-term preferences to encourage discovery beyond familiar shows. Both are collaborative efforts to connect a listener’s intent with concrete content to hear.

Raimond also describes generative personalization extending into the listening material itself. In his presentation with Wood, he showed an announced personal-podcast experience that could generate a recurring audio brief about a listener’s community. This extends his focus on listener control from choosing existing content to shaping the content and experience around what someone wants to follow.