Platforms like Spotify use AI and other machine learning tools to learn what you like. Nowadays, we affectionately call it “the algorithm,” but what is it really, and how does it work? While the specifics are kept under lock and key at Spotify HQ, we do know how the algorithm works in broad strokes. If you’re curious or looking to optimise your recommendations, read on.
The Basics
The first thing you should know is that most of the work happens the moment a song is introduced into the Spotify ecosystem. That means, by the time you play a song, the algorithm has already ingested it, analysed it, and theorised on the many different songs you’d want to hear next.
The first thing it looks at is the metadata. This is what you’d expect – the name of the song, the album, the artist, and other pertinent details, including background songwriters and producers. You may have noticed that, if you listen to a lot of songs with a word like ‘fire’ in the title, it shows up in the name of a recommended song too. That’s a simple 1:1 word association.
Genre is the other half of basic algorithmic matching. It helps in recommending new songs (but it also goes a little deeper than that). For now, at the basic level, listening to a lot of one genre will make more of its songs show up in your recommendations. It’s not much different from how YouTube and Netflix’s recommendation systems work, using common genres to anticipate what you’ll want to see next. Spotify figures out genre in two ways – it ‘listens’ to music and pays attention to the tags on your songs. There’s a reason a lot of services, including Spotify, give you toggleable filter tags for genre; it helps feed the algorithm.
The tags are also just popular. Sites with no need for predictive AI algorithms, like online casinos, still use filters as a way to let players skip to the good stuff. If you’re at an online casino with themed games, you’ll likely see filters for things like Christmas or genre-based brands like the Big Bass fishing slot series.
How Spotify ‘Listens’ to Music
The idea of Spotify listening to music may be amusing, but it is technically possible by parsing through raw audio. We can’t (and shouldn’t) try to comb through audio to quantify it, but fast-thinking machine-learning algorithms can and do. They establish key metrics that they can use to score a song. How many? We don’t actually know, but we can infer a few of them.
Energy would be an obvious one. How energetic is the track? Some are intense, some are lethargic. Similarly, every song has a more positive or negative sound, emotionally speaking. Machine learning can learn what sad or angry lyrics sound like, and how they differ from happy, bouncy lyrics. By scoring along these metrics, the algorithm can avoid recommending upbeat, bombastic tunes when you prefer slow, plodding, low-tempo music.
More analysis follows, looking at track lyrics for genre-associated keywords, visual cues from cover art/artist profile picture, and even a quick look at the internet to see how a song is being covered by pundits and fans. That last part happens thanks to OpenAI, whose internet-enabled LLMs form the core of Spotify’s modern algorithms.
Spotify’s AI-powered transformation has reached its zenith with Prompted Playlists, allowing users to summon playlists using their words. Naturally, the LLM uses those words to draw mood and usage inferences, finding music that invokes the same themes.
All of this information is then compiled and saved to your account, so the algorithm isn’t working overtime to keep up with you. Instead, it has a pool of information that includes your most played, manually saved songs/artists, and even demographic and location information. That, and the deep dive into the soundwaves of your favourite songs, makes it much easier to recommend something you’d bop your head to.