Problems in the Music Industry: The Need for Better Recommenders and Improved Music Discovery

Posted on Aug 6, 2020


Since artist compensation is dependent on the number of streams an artist can generate and the barrier of entry into music, in terms of distribution, is much lower, the entire pool of music available for streaming has become oversaturated with subpar songs and, more generally, the amount of readily streamable music is increasing rapidly. Building improved music discovery services and recommendation systems would solve these issues.

Artist Compensation and Counting Streams

Firstly, the current model for compensating artists (adopted by Spotify and Apple Music, the two largest music streaming service providers) is entirely dependent on the number of streams an artist can generate. Therefore, in the interest of making more of their product, artists have started to put out albums with high numbers of individual tracks. “Releases by Migos (Culture II, 24 tracks), Rae Sremmurd (SR3MM, 27 tracks), and Drake (Views, 20 tracks; More Life, 22; Scorpion, 25) don’t encourage being heard in full on one sitting, in fact, they actively discourage repeat listens. They’re intended to be cherry-picked for playlist consumption because it results in bigger streams and therefore higher chart positions.” (Triple-J)

Having to listen through 20+ track albums can be extremely cumbersome to the user, even if it’s their favorite artist. But while the current trend of pro-rata compensation continues the length of albums isn’t likely to decrease. Artists are incentivized to either shorten songs in order to cram more individual tracks into an album or they will become more liberal with the definition of an album-worthy track. Both result in much lengthier, high track volume albums. A solution to the overly long album problem is to provide a service that can quickly play through all the songs on the album, hastening the process of the user “cherry-picking” the tracks they like the most. This would take the burden of manually skipping junk songs off of the user, especially since most people know within the first 10-30 seconds whether they like a song or not (Philip Marsden + Tiny Room). This method of music discovery, where new songs are fed to you every 10-30 seconds, could also provide useful data to artists as it helps determine which songs are the most grabby in a given set.

A more institutional solution would be to compensate artists based on the minutes listened. And in some sense, the existing use of a threshold to dictate how long a user must listen to be counted as a “stream” helps disincentivize putting in songs that may not grab the listener’s attention. But while a pro-rata system exists for compensation, artists will always be incentivized to generate as many streams as possible.

Additionally, the speed with which artists are releasing music is increasing, due to their increased reliance on streaming revenue. This is the result of revenue from touring is near zero for the foreseeable future, CD sales are down, and merchandising is difficult without tours. Daniel Ek said that (“artists would no longer be able to release music every 3-4 years.”)[] Therefore, there will be more music to go through for the average listener interested in finding new music. A service that helps quickly cycle through new music while learning from your session tastes would help solve this issue.

The Barrier to Entry in Distribution

Another cause of the streamable music pool being oversaturated with bad music is that the barrier to entry has been lowered, particularly in terms of being able to distribute one’s own music. With the introduction of services like SoundCloud and YouTube which allow any user to upload content, the amount of available music (or audio content in general) has increased drastically. Furthermore, digital distribution tools like Distrokid make it possible for an individual to distribute their audio content on multiple streaming platforms in an instant. There’s no longer the barrier posed by record labels and large distributors like Universal, Sony, and Warner. This has led to many songs being made available for streaming that have never had to pass through any filter other than the artist’s own ears. The digital distribution of music has also made it very easy to reach a wide audience. Whereas a couple of decades ago you would have to hand out your mixtape on the side of the street and only hit a very local market, now you can make your music available to practically anyone in the world with very little cost.

The issue resulting from a lower barrier to sharing music with the world, that there is more subpar music diluting the total pool of streamable music, can be solved by improving recommender systems for music. Practical recommender systems have treated this issue by putting more weight on collaborative filtering methods, largely recommending music that other users have already listened to. While this does work well in terms of filtering out the songs no one should ever be subjected to, it also makes it much harder to discover brand new music. Furthermore, users often get siloed into a particular subset of music–not necessarily genre-specific but all related in some way–further exacerbating the difficulty of finding new, fresh music.

Summary + Solution

Ultimately, there is an oversaturation of songs most users will ignore in the total pool of streamable music, but also just a surplus of songs in general. This is the result of the models of artist compensation large streaming services abide by, where artists are paid based on the portion of total streams they generated. It’s also the result of the new ease with which amateur artists can share their self-made music through digital distribution services and platforms like SoundCloud and YouTube.

As touched on earlier, there are a couple of approaches to treating the issue of oversaturation in music streaming, some already in practice but in need of improvement. The first solution is in response to the growing trend of lengthy, high volume albums and quicker release rates. Artists are incentivized to pack as many individual tracks as possible into a package due to their returns being dependent on the number of streams they generate. More importantly, the rate at which artists release is going to increase drastically, especially with new or rising artists. This is due to streaming revenue making up a larger portion of artists’ total revenue, especially with touring at a standstill. This means a service that feeds new music to a user over a short duration (i.e. starts at 10-30 seconds) and actively learns from the user’s session activity (such as liking or skipping songs) would be very useful. It will hasten the process of discovering new music, and can further provide the artist and service provider (me) with useful information concerning the listener’s habits–which songs stick and how long before users skip (if less than set skip rate).

The second solution is a response to the oversaturation of streamable music with subpar music as a result of the lower barrier to entry in terms of distributing music digitally. The simple treatment is to have good, robust recommender systems that deliver content the user engages with–whether that’s repeat playbacks or listening for more than 30 seconds. While practical recommenders are quite powerful and robust (Spotify’s is world-class, and it’s why I pay them and not anyone else on the market), they often put more weight on collaborative filtering. This leads to the typical recommendation being a song someone else has already listened to and increases the difficulty of finding brand new music. Furthermore, the typical Spotify user thinks in terms of playlists which are representations of some idea, mood, or feeling of the user. When they are recommended new music and they enjoy a song, it also encompasses some mood or context. All the songs they like within a certain time period are related to each other in some way, much like a playlist. Therefore, taking a session-based approach to music discovery, where a user provides some seed track or playlist and is fed new music in short intervals based on their likes/skips in the current session, would be useful. The service would leverage existing recommendation systems but would learn based on user activity within that session. Because sessions would take place within one fixed context–both environmental and user-specific–this would provide valuable data for the purpose of building a context-aware recommender. Then, music could be automatically pushed, with nice fluid transitions, based on changes in context.