Spotify, a music streaming app with 182 million subscribers and 422 million users that use the app at least once a month, is no stranger to the power of AI and data. Spotify is looking to change the music industry with AI generated songs, but this is not the only way the company is incorporating AI as it seeks to differentiate from other music streaming services. As a surprise to me, AI use in the music industry dates back to 2017! Sony created a hit song “Daddy’s Car” in 2017 which has to date 2.9 million views on YouTube alone. The song, created through a program called “Flow Machine” created a new original sound with features akin to the Beatles as programmed in the algorithm. Following the apparent success of this song, Spotify poached the Sony artist and producer who worked in Sony’s Computer Science Lab to head Spotify’s Creator Technology Research Lab division, its own internal AI lab. The purpose of this division, established in 2017, was to focus on making tools to help artists in their creative process.
AI through Acquisitions
Spotify gathers millions of data points on its users from which songs are streamed by users, which ones are saved, which artists/songs/albums are searched by users, etc. This data collection enables the company’s recommendation features which have over time become even more accurate reflecting user preferences. While AI’s applicability in the field of music was still nascent in the mid-2010s, companies such as Google (Magenta) and IBM (Watson) were looking into how to teach computers to recreate human creativity.
Spotify, not one to stay behind in the streaming game, has pursued a number of acquisitions since 2013 to beef up its AI capabilities. In 2013, it acquired Echo Nest bringing in-house expertise in acoustic analysis and machine learning. This led to the music curation features through personalized playlists such as Release Radar, Discover Weekly, and Daily Mix, leveraging the extensive data the company was already collecting on its millions of users. Additionally, Spotify acquired in 2017 a startup called Niland, a music technology company that provides music search and discovery engines based on deep learning & machine listening algorithms.
The integration of these acquisitions helped Spotify leverage three kinds of AI models to produce robust personalized playlists. These playlists create value for the company through increased number of streams as well as visibility to new artists by exposing them to users that might have similar music preferences to the content they produce.
AI Models Used by Spotify
1. Collaborative filtering (comparing user preferences like Netflix’s algorithms does through ratings but instead through how many users save songs, number of times a song is played or how many times users click to the artist page)
2. Natural language processing (AI scans a song’s metadata and blogposts about musicians, articles on the internet and social media to see their relevance), and
3. Audio models (analyze and categorize songs)
Evolution of AI Music and Creating Tools for Creators
Flow Machine, developed by Sony, is an AI system, that works by first analyzing a database of songs, and then following a particular musical style to create similar compositions. However, the final composition still has a touch of human skill in the arrangement of the song and the writing of the lyrics. This was the program used to create the “Daddy’s Car” song and proof that AI-generated music could in fact become a hit song, regardless of the lack of association with a known artist.
Spotify dived in head straight into the world of AI bringing in over the Sony veteran to lead its very own in-house research lab. This move likely reflected the company’s willingness to explore how AI can enhance value creation for collaborators, artists and more, beyond helping Spotify fine tune its playlist recommendations. Additionally, it means the company understood the need to have a dedicated division within the company to explore different use cases for AI and run different experiments before bringing them online into the app.
Spotify has its own online music-making studio, Soundtrap, and in 2022, the company launched an open-source AI-powered tool called Basic Pitch to help users upload audio recordings and help them convert it into MIDI (musical digital standard used by electronic musical instruments). In 2022, Benoit Carré (AI-generated music veteran) released his new album Melancholia, an 11-track collection created using the tools developed by Spotify’s lab.
Spotify faced controversy in the music community and press after it was found many of its playlists included songs associated with fake artists that were traced back to the Swedish company. A big issue regarding AI-generated music goes back to the royalties and ability to protect intellectual property. After all, if the AI algorithm created the underlying melody, is it copyrightable? Can the artist whose “similar” melody was fed into the algorithm able to argue in a lawsuit that it deserves royalties related to those streams?
Many have criticized Spotify’s venture into AI-generated music as a way to circumvent royalties by generating songs which it would own the rights to. With the amount of data the app gathers on users, it can easily use the AI tools it is developing to create songs that would be instant hits and perhaps employ a couple of in-house songwriters to write compelling enough lyrics to accompany the AI-melodies.
Spotify is now also looking to encourage regular users to access AI tools to create their own music. For other platforms such as TikTok, Facebook and Instagram, user generated content has been key to their growth. Questions remain whether this is really going to create additional value for Spotify’s user base and the company itself.
Challenges remain in the road ahead for Spotify’s development of new use cases for AI and ML to enhance its offerings to users and artists. However, the biggest question is how Spotify will navigate the consequences of implementing these use cases vis a vis music studios and artists’ interests and the implications regarding copyright when AI models are trained on existing songs and melodies and asked to produce something similar to something that already exists.