Long gone are the days where music was physically bought and played on a disc player and had to manually be changed in order to listen to another album. We have seen a slight comeback in the vinyl industry, but frankly it’s more for the aesthetic purpose rather than the convenience us innovative and lazy humans strive to achieve.
With the emergence of the digital music library came a method for people to pick and choose the songs they bought and create a playlist that they rapidly abused, never to listen to those songs again (sorry if it’s just me). Soon, even the process of buying individual songs and albums became a thing of the past and now a new type of method of music listening is king: subscription based music streaming. For the price of a brand new album a month ($10), users can have access to tens of millions of songs and albums to listen to, all ad free, and all online or downloaded depending on what you want.
As platforms like Spotify and Apple Music began to develop their services more, along came the process of music curation. Depending on the genre, a group of experts would use data and the latest trends to create a playlist that they deemed the most trendy for that genre. Although this approach was good for some time, there was a problem with this: not everyone liked every song on it.
The answer to this problem? Machine Learning.
Spotify’s ‘Discover Weekly’ is one of the best examples of the integration of machine learning in music streaming. Each week, users get a playlist of thirty songs that Spotify thinks the user will enjoy. This list incorporates songs the user may not have heard in the past; however, the list was comprised using a multitude of machine learning algorithms that we will explore further.
Collaborative Filtering: This form of AI works just like the Netflix recommendation system you may be familiar with. After watching a movie, Netflix may ask you to rate it. Based on the rating you give the movie, Netflix is able to recommend movies and TV shows that you may like based on what others rated the movie you just watched and also liked. Since songs cannot be rated as easily as movies, Spotify uses a different metric for collaborative filtering: play counts. Depending on how much you play a specific song, Spotify can safely assume that you liked this song and can then recommend songs that others have listened to with similar preferences. In a sense, this algorithm would not be possible without a community to collect and interpret data from.
Natural Language Processing: Like the name may suggest, this method of AI uses human language to recommend songs to people. It does this by going through the internet to look for buzzwords that may be relevant to certain musicians or songs. The top words that come as a result are piled into what Spotify calls “top terms.” With each artist or song having thousands of terms that change on a daily basis, this process is ongoing. Each word, however, has a strength associated with it that determines how likely a user is to classify that song with that word. Based on this strength, Spotify can recommend songs that it finds fit according to your preferences.
Raw Audio Models: The last of the bunch, raw audio models are very useful when it comes to songs that are relatively new. With collaborative filtering and natural language processing, a lot of data is required to be able to recommend a song, but a new song, for example, may not have lots of coverage on the internet nor have the same number of streams as older songs. This does not mean that it is a bad song, so raw audio models essentially look at the song for what it is: audio waves. This very complex process is done through convolutional neural networks. In its more basic definition, neural networks act like a human brain and can process information and learn from it to draw its own conclusions about something. In the same sense, raw audio models can give the program an idea of the characteristics the song has. For example, it can see if a song is high or low tempo, or can even recognize metrics associated with certain genres. Then, by utilizing the other 2 forms of recommendation, Spotify can deliver a solid choice to listen to.