Three Chords and the Truth (with a little help from AI)
For me, music has always been a huge part of my life with my Spotify account constantly playing in the background satisfying my love of country music, 90s cheese classics, and pop punk (a fairly eclectic mix!). As Artificial Intelligence vastly moves away from the notion of building robots and more into how it’s incorporated into our daily lives, I thought it would be interesting to look at how AI is currently being adopted in the music industry.
I have always seen music as highly personal and subjective, so I was fairly surprised to see a variety of articles and analytics case studies online combining music and AI. A key area where AI appears to excel in the music industry is via personalised music recommendations. A ‘key’ influencer in this trend has been Spotify with their Discover Weekly Playlist which provides listeners with 30 tailored new song recommendations each week. Personally, I have found quite a variety of new songs and artists which fit my unique taste, making me wonder can AI really help predict something as subjective as music?
Songaz (available in the US & Canada) opened the gateway to music recommendations in the early 2000s by employing a team of ‘music experts’ to make playlists. These ‘Experts’ would produce playlists linking to categories such as time of day/mood/activity i.e. ‘Waking Up’, ‘Unwinding’, ‘Working Out’. Each of these songs were manually picked by the experts and put into these fairly generic categories. Pandora later took this one step further, by ‘manually tagging attributes’ whereby groups of listeners gave descriptive words for songs attributing them with a ‘tag’ which could then be grouped into playlists of songs with similar ‘tags’.
Spotting your next favourite artists
Today Spotify’s Discover Weekly Playlist elevates recommendations to the next level using three main types of recommendation models to ‘treble’ the accuracy of personalisation.
Spotify uses collaborative filling via implicit feedback obtaining data such as stream count, saved songs and viewed artists. Once this information is collected, collaborative filling then finds similar users based on their taste in tracks / artists, by running them through a Matrix Maths via Python Libraries. This ultimately leads to recommendations based on what similar users are listening to, that isn’t in your music library (yet!).
Alongside Collaborative Filling, Spotify also uses Natural Language Processing (NLP) to crawl through the internet, reading blogs and articles to figure out what people’s opinions on songs and artists. NLP analyses the particular language used to describe the music whilst also processing other artists referenced in the literature. This data can then be used to identify similar pieces of music which can again aid recommendations.
The final method Spotify uses to ‘tune’ into the perfect recommendation is Raw Audio Models. The above models can generate solid personalised recommendations, however, they fail to account for new music which is pretty integral considering the rapid ‘tempo’ of the music industry. Raw Audio Models don’t give weight to the number of streams/likes of songs which allows for new song and smaller (new) artists to be included in Weekly Discover Playlists. Raw Audio Models work similar to Facial Recognition Software, however, analysing audio data rather than pixels! Raw Audio Models analyse factors such as time signatures, mode, tempo, volume and key to link musical similarities between songs/artists to generate personalised recommendations using individual user’s listening library.
These models are then brought together in Spotify’s network to bring us our own personalised weekly playlists where you may discover your next favourite artist!
Compose yourself Mozart, AI is coming
The notion of using AI to derive personalised recommendations does begin to remove some of my preconceptions around the subjectivity of music however, can this be pushed one step further with AI predicting/generating hit songs?
A study by Herremans et al. (2014) created a model to predict if dance songs would be in the top 10 dance chart. This model analysed audio features and managed to produce a model which had an AUC of 81%. A later study which was carried out in 2017 by Herremans & Bergmans combined both audio and social listening data to predict if a song would be in the top 20 dance chart with 79% accuracy. These studies suggest that chart prediction is possible and with AI models continuing to learn and as data sets expand it is likely these models will ‘amplify’ their accuracy in the near future.
AI is now taking a step even further in the music industry as it attempts to compose music! An example of this is from YouTuber Taryn Southern who has used AI to bridge the gap in her lack of knowledge for music theory. Last year she released her latest album 'I am AI’ for which is completely composed by four AI programmes: Amper Music, IBM’s Watson Beat, Google’s Magenta, and AIVA. Southern’s AI based music career has achieved her 13.4K monthly listeners on Spotify and could signal some traction for the future of AI generated music.
This year has also seen the launch of a new app called Mubert which uses algorithms to produce original electronic music in real time personalised to your current mood with options such as study, relax, dream etc. This new app generates endless streams of beats across six genres; deep house, trap, chill step, trance, ambient and liquid funk with users given the option to modify aspects of their sound further e.g. by tailoring the intensity.
From Chico Time to AI Time?
AI is becoming increasingly involved in the music industry with personalised recommendations successfully bringing new music to users and providing leverage and insights to new artists. I am still sceptical on the success of AI in terms of composing ‘groundbreaking songs’ and ‘replacing’ artists, however, the music industry is privy to music phenomenons with Crazy Frog, Bob the Builder, Teletubbies and Chico Time all having taken the top spot in the charts!