Between complex algorithms, behavioural data and artificial intelligence, music streaming services sculpt our daily listening. But how do their recommendations really work? A dive into the invisible machinery that influences our playlists, our discoveries… and perhaps even our emotions. Every day, millions of listeners discover a new “recommended” track on their
Every day, millions of listeners discover a new “recommended” track on their favourite streaming service. And yet few know what goes on behind the scenes. How do these platforms pick the track that will slip into your current mood, your evening or your morning run? The answer lies in a complex balance between data science, advanced music analysis, machine learning and algorithmic engagement strategy.
One of the foundations of recommendation systems is collaborative filtering. This mechanism, now a classic, compares your behaviour with that of other users who share similar tastes. If thousands of people who listen to Bicep also add Overmono to their playlists, the likelihood of Overmono being suggested to you goes up. Spotify, for example, applies matrix factorization techniques (via ALS or SVD) to project users and tracks into a shared vector space, where proximity becomes a recommendation criterion. This approach has proved powerful, but it still depends on the density of listening data.
To get around the limits that come with new releases or little-heard tracks, platforms add a detailed analysis of the audio content. This method, known as content-based, relies on automatically reading the sonic DNA of tracks: tempo, key, energy level, danceability, duration, rhythmic structure, and so on. Thanks to machine learning techniques, such as convolutional networks applied to spectrograms, the algorithm can match two tracks based solely on their sonic texture, even without any user interaction. Deezer and Apple Music make heavy use of this layer, cross-referencing it with human tags (genre, mood, atmosphere) to fine-tune suggestions.
But the current trend goes further, introducing sequential models. The order you listen in, the time of day, the day of the week, the context of use: everything comes into play. For several years, Spotify has been testing RNN-type architectures (or, more recently, transformers) that model not only what you listen to, but how you get there. The platform can then predict a natural sequence of tracks, as a DJ would, drawing on the listening behaviour of entire sessions rather than isolated events. This is the very logic behind Discover Weekly, Release Radar and the Daily Mixes.
Artificial intelligence, meanwhile, is now built into every layer of the process: automatic recognition of emotions in the voice, generation of fictional listener profiles to test recommendations, real-time adaptation to your reactions, and soon, perhaps, the creation of playlists entirely driven by AI agents tailored to your intentions, rather than just your habits. YouTube Music, for example, already uses video data to adapt its recommendations in a multimodal way, learning from visual content too. Apple Music is starting to experiment with explainable systems, able to tell you why a particular track was suggested to you: a promising avenue for building trust and engagement.
For artists and labels, these systems aren’t just a black box: they can be influenced. A good completion rate in the first 48 hours, additions to personal playlists, a low skip rate or few early skips are all positive signals sent to the algorithm. The challenge is no longer just to make a good track, but to release it in conditions that will allow it to break through.
Tomorrow, AI could go even further: recommendations based on your emotional state as detected by your connected devices, live adaptation of a setlist to your physiological reactions, or even tracks generated or modified on the fly according to your deepest preferences. This is no longer just the future of streaming: it’s the future of our entire relationship with listening to music. And it is already being written.
