Your Next Jam: The 3 AI Models Behind Spotify's "Discover Weekly" Playlist

Released every Monday, the Discover Weekly playlist feels like a personal DJ service—and that's exactly what it is, powered by three distinct AI models working in harmony.

Spotify's **Discover Weekly** is arguably the most successful application of music recommendation AI in the world. Its secret sauce isn't just one algorithm, but a powerful triad of Machine Learning models that cross-reference user data, textual descriptions, and the raw sound of the music itself.

1. Collaborative Filtering (The People Model)

This is the most common recommendation system. It answers the question: **"What are other users like you listening to?"**

• **How it works:** If you and another user listen to 90% of the same artists, the model assumes you'll enjoy the one song they listen to that you haven't discovered yet.
• **Focus:** Matching users with similar listening histories. This is the foundation of the playlist.

2. Natural Language Processing (NLP) (The Text Model)

NLP helps the AI understand music through external context—the words used to describe it online.

• **How it works:** Spotify scrapes the web (blogs, music reviews, social media) to find terms associated with a song or artist. If your favorite bands are frequently described as "lo-fi chillwave," the NLP model will suggest other bands tagged with similar vocabulary.
• **Focus:** Understanding genre, mood, and cultural context. This model helps find entirely new artists that fit your *vibe*.

3. Raw Audio Models (The Sound Model)

This is the most advanced layer. It ignores human labels and analyzes the music's waveform itself.

• **How it works:** A Convolutional Neural Network (CNN) listens to the song, analyzing musical elements like tempo, key, acoustic qualities, volume, and instrumentation. It can identify a track's "sound" fingerprint and match it to other tracks that are sonically similar, even if they're completely unknown or untagged.
• **Focus:** Identifying deep sonic similarities and filling in the gaps where data is sparse.

The Discover Weekly playlist is the output of these three systems arguing (and agreeing) about what you’ll love next.

Every time you skip a track, you’re instantly providing feedback that fine-tunes all three models for next Monday!