Ever since Spotify launched in South Africa back in 2018, the music streaming platform has seemingly been going from strength to strength.
Part of the success has been the new avenue for local artists to be discovered and while there is still some conjecture over how big a cut they receive, the other major contributor for its rise over competitors is Spotify’s algorithms.
It too is shrouded in mystery, but whatever it’s make up, the recommendations it helps serve up and personalisation it delivers, is market leading at the moment.
In order to gain a bit more insight into how Spotify’s algorithms keep people listening, we recently spoke to Ziad Sultan, who is the senior director of Product for Personalisation at the company.
He does not give us all the secrets to Spotify’s algorithms, but does unpack how large the team working on it is, as well as some of the machine learning and other tools that the platform has in place to deliver said personalisation.
Here’s what he was able to share.
Hypertext: Spotify’s personalisation is the audio streaming equivalent to the Coca-Cola recipe or KFC’s 11 herbs and spices. How many people work on it and can you tell us more about what makes your approach so unique?
Ziad Sultan: At Spotify, we have more than 7 000 employees working on providing the best audio streaming experience. Our personalisation team is made up of hundreds of employees focused on bringing listeners the right content at the right time.
There isn’t just one experience but more like 365 million – one for each Spotify user.
Through the use of machine learning and some of the most powerful algorithms, we’re essentially able to provide users with an ever-evolving soundtrack to their life, by continuously listening, learning and leveraging their habits and interests.
How does personalisation at Spotify work? How long on average does it take for your algorithms to attune to a user and what are some of the metrics it focuses on?
ZS: Once you’ve signed up to Spotify, our system immediately starts picking up your interests and listening behaviours.
We’re able to make great personalised recommendations because of complex, dynamic systems that consider a wide variety of inputs about what you like—which we refer to as signals—and balance those in many possible different pathways to produce an output: the perfect song for the moment, just for you.
Our system picks up thousands of types of signals: what you’re listening to and when, which songs you’re adding to your playlists, the listening habits of people who have similar tastes, and much more.
Does Spotify analyse every single song on the platform for BPM, genre, etc. and is this process automatic or manual?
ZS: Historically we’ve relied on user categorisations of the music they listened to. When we introduced Spotify in 2008, the consumer software that was launched brought access to a catalogue of music and we asked the users to organise this with us. Listeners started creating their own, highly personalised playlists representing their moods, mindsets, tastes or habits.
In that, they essentially organised the world’s music catalogue in all types of dimensions, labelling the tracks for the rest of the system and the Spotify ecosystem. They became mirrors, providing rich and textured data sets, which were personal and uniquely reflected the value and personal use cases of the music they engaged with.
Since then, we’ve introduced a whole range of additional systems that automatically pick up more insights about the songs on Spotify such as speed, length and musical elements of songs to power the recommendations that our users know and love.
Today we capture more than half a trillion “events” on our service every day from listeners interacting with music and podcasts on Spotify.
Do algorithms change according to the country that the user is from? And do local artists get any kind of advantage from algorithms in their home country?
ZS: Today Spotify drives 16 billion artist discoveries every month and the algorithms that drive these discoveries rely on a whole range of situational signals, including location, language, time of day, interests and many more.
As we continue to expand our personalisation efforts, we know that relevance is key to ensuring the best possible listening experience, which matters to both listeners and artists.
On top of that, we also use a unique approach which we call “algotorial”. To avoid overly relying on past data and potentially missing cultural moments before they become popular, we leverage an elite team of editorial playlist curators to teach our machine learning system. They create playlist experiences with thousands of tracks that fit one of their intuitions for how people might consume music in different situations.
As a result, our recommendations reflect cultural trends as they’re happening, elevating great content before it breaks and introducing new opportunities for artist discovery.
While end users experience the impact of algorithms in the form of better recommendations, does it hold any value for artists? Does Spotify share any insights on that front?
ZS: Better recommendations mean more opportunities for artists to be discovered and grow their fanbase. On top of that, we also make a whole range of features available to them, like profiling tools where fans can showcase their artistry and share updates, or analytics tools where they can keep track of their performance.
Artists can also pitch their songs to certain playlists. There are several factors such as mood, instrumentation, ambience, and tempo that can help music editors potentially find an ideal spot for an artist in our playlist ecosystem – a spot that not only puts you on a playlist but puts you in front of the right audience.
More creators are creating — and succeeding — than ever before. In 2020, over 76,000 artists were added to Spotify playlists for the first time, the large majority of which were discovered through our playlist pitching tool, which is freely available to all artists.
Also, as of 2020, 57 000 artists now represent 90 percent of monthly streams on Spotify — a number that has quadrupled in just six years. That means a growing and increasingly diverse group of artists are finding fanbases.
How often are your algorithms refined and what would it take for Spotify to integrate new information?
ZS: We continuously refine our algorithms to enable even more fan discoveries of new artists and offer the best possible unique listening experiences.
The Spotify we know compared to when the service first launched in SA a few years ago is quite different. How did the experience evolve in that time?
ZS: The most obvious way Spotify has dramatically changed is through the availability of new personalised playlists.
Since the day we first introduced Discover Weekly, there’s been a whole range of new playlists we’ve made available to listeners worldwide, including Daily Mixes, Release Radar, Time Capsule or Blend, each powered by our state-of-the-art machine learning expertise to deliver you the right content at the right time or occasion.