Skip to Main Content

How Spotify is using machine learning to offer you music you’ll love

Posted by Charlie O'Toole on 16th August 2019

Technology has long been a part of the music industry. When multi-track recording came into force in the 1950s, it was the musical equivalent of the moon landing. Before that, if you wanted to record a song, you had to do it all in one take with every instrument that you wanted on the track.

Then in 1966 Arbiter Electronics released the ‘Fuzz Face’, a key component to Jimi Hendrix’s sound, which recreated the distortion originally found in busted speakers. The concept of having pedals that could alter the sound of your instrument was not new, but Hendrix and Keith Richards bringing these ‘stomp-boxes’ into the mainstream revolutionised the way that music was played.

Now in 2019, machine learning is revolutionising the way that we find music.

Anyone who has used Spotify’s ‘Discover Weekly’ playlist has seen how eerily accurate it can be in suggesting new music to you. This is no accident; Spotify has invested heavily in its machine learning capabilities as part of its strategy to assert market dominance over heavy competition from companies such as Pandora, Apple Music, Tidal, SoundCloud, Amazon and Google. Spotify’s IPO prospectus said as much with it promising to “continue to invest in our artificial intelligence and machine learning capabilities to deepen the personalised experience that we offer to all of our users”, and that “this personalised experience is a key competitive advantage.”

To achieve this, Spotify employs three types of machine learning: collaborative filtering, natural language processing (NLP), and raw audio models. Collaborative filtering involves recommendations based on those with similar tastes to yourself. With NLP, Spotify analyses a plethora of data sets from blogs to articles, band profiles, song metadata and more to find which artists are commonly mentioned alongside each other.

Raw audio models then are the pièce de résistance of Spotify’s machine learning strategy. While you may be aware of other artists in a similar genre, or may know who has collaborated with who, this is the part of the system that allows Spotify to throw you a curveball that you may never have thought of but end up loving. By analysing the musical elements of a song (tempo, time signature, key, etc.), Spotify is able to recommend you songs that have similar features to what you listen to most from artists who aren’t in any way associated with those you already listen to.

This is fundamentally changing the way that we find music and the way that artists are discovered. A recent example of an artist who was brought into mainstream consciousness through Spotify’s algorithm is Dermot Kennedy. One of his songs ‘Glory’ was picked up by the algorithm and was placed on users Discover Weekly, and soon after his songs were performing so well that the CEO of Spotify, Daniel Ek, was notified. Kennedy himself recently emailed Ek to thank him for all that his platform has done for him. The success his EP saw from streaming allowed Kennedy to make a living purely from royalties and gave him great leverage when it came time to sign with a label.

However, the attitude towards streaming platforms is not always as positive as my colleague Mark recently mentioned in his blog post. Many legacy artists are resisting the urge to transition to streaming platforms due to the low royalty rates (reportedly between $0.006 to $0.0084 per stream on Spotify). This contentiousness may be due to the fact that revenue from streaming is only a fraction of what the revenue from physical sales used to be, and these artists are reluctant to embrace the changing landscape of the music industry.

With the age of DJ’s dictating musical taste to the masses coming to an end, all that’s left to wonder is who will be the next artist to be crowned by Spotify’s algorithms.