`2 When Do Files Become Music?`

“Spotify paints it black.” This message appeared on the Spotify company blog in January 2015, with the promise to bring phone users “the best-looking Spotify ever.” The blog post claimed that, with the darker theme, playing your favorite music has “never looked so good.” With its “refined interface,” the new look “lets the content come forward and ‘pop,’ just like in a cinema when you dim the lights.”¹

Interfaces indeed pop forward—and, by doing so, hide their infrastructures. “When we purchase a device,” anthropologist Lane DeNicola once noted in regard to the iPhone, “we are purchasing access to its ‘surfaces,’ a transient experience of use divorced from either internal mechanism or the particulars of production.”² Transparent devices are highly intuitive, constituting an interface that rapidly disappears from our perception. The same holds true for a service such as Spotify, whose software and content are patented and copyrighted, whose standards and protocols are hidden, and whose graphical user interfaces are designed to make invisible the overcoming of any remaining barriers to downloading or file sharing, such as social interaction or technical skill.

This is not to say that listeners are unaware that experiencing music as software differs from listening to a CD or an LP. The “lean back” experience is less prominent. Online input is constantly needed. Active listeners are familiar with the prescriptive demands of the service and the ways that Spotify summons its users: “Know what you want to listen to? Just search and hit play. Go get the music. Check out, discover new tracks, and build the perfect collection.” As Jeremy Wade Morris argues, music as software has introduced a new “technological relationship” to processes of search and discovery, listening and liking, exchanging or buying music. When streaming services code music and redefine it as a data-driven communicative form—with audio files and metadata being aggregated through various external intermediaries and with user-generated data being extracted from listening habits—the singularity of the music experience is transformed into what Morris has termed “a multimediated computing experience.”³

Yet participating in multimediated computing experiences does not necessarily entail an awareness of the broader digital communication infrastructure of which Spotify is a part or of the sociopolitical environment in which this communication originated and to which it is addressed. In order to understand the logic and rationale of streaming services such as Spotify, we need to ask what exactly happens when data are turned into music and vice versa. The focus of this second chapter, therefore, is to find out what happens beneath the shiny black and green surface of Spotify’s interfaces, or how music streaming is infrastructured. We speak of digital media infrastructure in a broad, inclusive sense of material and sociotechnical forms that allow for the possibility of exchange over space, and we understand infrastructuring as a process whereby this exchange is initiated, managed, and observed.⁴ While the first chapter has followed the hype and mapped it over an investment timeline, the second chapter thus peels off a further layer, as it discusses Spotify’s diversified data infrastructures and critically examines what happens prior to and during the computational experience of “listening” to files becoming music. After sketching out the contours of this process based on publicly accessible information gathered via trade journals and company blogs, among other sources, the chapter concludes with a first experimental case study.

`Following the Data`

Research on the cultural implications of software—whether in the form of software studies, digital humanities, platform studies, or media archaeology—has repeatedly stressed the need for in-depth investigations of how computing technologies work, combined with meticulous descriptions of technical specificities. Our analyses of Spotify relate to such an interest in the specifics of the computational base—that is, the mathematical structures underlying programs and interfaces—and in technically specific ways of understanding the operations of material technologies.⁵ Yet what is at stake is not a complete understanding of all mathematical concepts and models driving these programs. The claim that studying culture through data necessitates “a thorough training in programming so as to allow researchers … to ‘look’ into the ‘black box’ of the technology … misses the mark,” as Mirko Tobias Schäfer and Karin van Es have argued.⁶ Rather, as we have previously described, our aim is to understand Spotify’s data infrastructures sufficiently to approach the software with novel research questions, critical discussions, and productive experiments.

As discussed in the introduction to this book, media environments today are “increasingly essential to our daily lives (infrastructures),” yet they are often “dominated by corporate entities (platforms).”⁷ The relation between these two concepts remains tricky, if not dialectical, with most scholars opting for either platforms or infrastructures as a critical lens. Platform studies have emphasized the dual nature of commercial platforms. YouTube, Facebook, and Twitter support innovation and creativity but also regulate and curb participation, with the ultimate goal of producing profit for platform owners.⁸ In short, platform affordances simultaneously promote and constrain expression. Spotify, however, differs from other platforms. It is a service that caters to record labels and artists by seeking to provide a regulated and commercialized streaming service with professional music, as opposed to an open platform with user-generated content. Given these differences between Spotify and SoundCloud or YouTube, the conceptual focus here will not be on Spotify as a platform but rather on the infrastructuring of this service and its environment: what technologies, data, and organizations have to come together to make Spotify happen?

Spotify’s proliferating data infrastructure includes more than meets the eye. Its “event delivery system” is one of the foundational pieces of that infrastructure. Most events that are produced within Spotify are directly generated from millions of Spotify clients as a direct response to certain user actions. All types of usage data can be defined as a set of structured events that are caused at some point in time as a reaction to some predefined activity. Whenever a user performs an action in the Spotify client—say, listening to a track or searching for an artist—a small piece of information (an event) is sent to Spotify’s servers. Event delivery, “the process of making sure that all events get transported safely from clients all over the world” to the central processing system, is consequently a key requirement of Spotify’s data infrastructure.⁹ But the latter is also dependent on actual music supplied from outside the service. If one opens the “About Spotify” tab, for example, it is strikingly apparent that music per se comes from elsewhere: “Content provided by” is followed by a series of logotypes for Universal Music Group, EMI, and Warner Music Group, among others.

As both Spotify researchers and listeners, we have repeatedly asked ourselves what types of data exchanges are mediated by the service and how these affect the ways in which music is distributed and consumed. The simple question of what a music or audio file is (or converts into) when being listened to has no simple answer, however. Data files become music on Spotify in various ways. Digital formats have made cultural commodities such as music “infinitely analyzable and quantifiable in ways that make them increasingly amenable to algorithmic exploitation.”¹⁰ Yet such numerical alterations at the back end remain invisible on the interface, or front end.

`Sharing Data`

Data exchange and interaction also occur constantly between Spotify and other companies, and it is these interactions that define what kind of service Spotify is. Spotify’s cooperation with the music intelligence company Echo Nest can serve as a case in point, since it is a partnership that indicates how datafied music consumption came to alter the company. In chapter 1, we charted the development of Spotify in detail. Suffice it to say, the Echo Nest was important for instigating Spotify’s data-driven music recommendation systems. Personalized music offerings were spearheaded by the American company Pandora Radio, however, which for a number of years was widely recognized as the best music recommendation service. Spotify’s partnership with the Echo Nest initially concerned a similar radio function, and in a blog post in December 2011, it was announced that the company would now power Spotify Radio: “Thanks to Spotify’s deal with The Echo Nest, users of the popular music service can now create streaming radio stations based on any artist or song on Spotify.”¹¹ The cooperation developed, and in April 2012, Spotify began updating its desktop software with several new features, including a Pandora-like radio station. An online radio offering, it was proclaimed, “would advance Spotify’s strategy of attracting users with free, ad-supported services who can be converted later into paying subscribers.”¹² At the time, the Echo Nest did not exclusively power Spotify’s radio recommendations. Since its application programming interface (API) was open, competitors such Rdio and Deezer were also using it. In March 2014, Spotify acquired the Echo Nest, a deal that was said to strengthen its music discovery expertise and to “allow Spotify to leverage The Echo Nest’s in-depth musical understanding and tools for curation to drive music discovery for millions of users around the globe.”¹³

In recent years, Spotify has put a number of employees from the Echo Nest in charge of its most important discovery products. While Spotify’s music discovery engine remains inaccessible to outside observers, Spotify itself also remains dependent on, if not conditioned by, actors outside of the service. As we have argued elsewhere, music distribution is increasingly driven by automated mechanisms that continuously capture, sort, and analyze large amounts of web-based data. Music metadata management and data ca(t)ching mechanisms powered by the Echo Nest have been key for the development of Spotify, with audio files constantly being linked and woven together with contextual information.

The partnership between Spotify and the Echo Nest thus suggests a more general observation regarding how files become music, as it makes it clear that audio files cannot be studied as isolated or autonomous entities. An audio file found and enjoyed via the internet is equally a product of different kinds of metadata. Metadata here need to be understood as a key element in setting the stage for what later becomes music listening, as they form a central part of contemporary algorithmic knowledge production. Consequently, metadata used by Spotify via the Echo Nest should not be perceived as an entity that sits discreetly in the background and assists those who actively search for it. Instead, metadata are constantly being marshaled to steer, direct, and reconceptualize both digital music artifacts and music recommendations.

On the one hand, these automatic systems for music information retrieval often reveal the diluted and flawed nature of online metadata. Confusion over names is a case in point. This is often not a major problem for popular artists, but for lesser-known musicians, poor metadata around name recognition can have devastating effects, leading to erroneous descriptions and recommendations. As noted elsewhere, our examination of the Echo Nest’s API and its data collection revealed that data concerning Danish pop singer Dorthe Kollo was inaccurate. Some data even pertained to different Dorthes, as the Echo Nest had confused Dorthe Kollo with both Danish author Dorthe Nors and Danish singer Dorthe Gerlach.¹⁴ Then again, the combination of the Echo Nest’s “music intelligence” technology—even if its sophistication can occasionally be disputed—and Spotify’s massive listening data trove has paved the way for popular algorithm-generated and personalized suggestions, such as the “Discover Weekly” playlist.¹⁵

The fusion of Spotify and the Echo Nest in 2014 thus testifies to what Anne Helmond has described as “platformization,” or the “extension of social media platforms into the rest of the web and their drive to make external web data ‘platform ready.’”¹⁶ As in the case of Spotify and the Echo Nest, such extensions do not come naturally; they are strategically driven forward and may sometimes fail. This also becomes obvious in regard to Spotify’s data integration with Facebook, an infrastructural tie-in that was highly visible across Spotify’s interfaces. The partnership with Facebook and related companies and actors is both instructive and explanatory, as it attests to the gradual transformation of Spotify from a tech company that simply distributes content produced by others into a globally operating media firm. As discussed in chapter 1, this cooperation is aimed at making the service more “social.” When Spotify merged its own login system with that of the social networking giant, it also caused a spike in new users. “Spotify introduces music to your social life,” Daniel Ek stated in a blog post in September 2011, referring to the F8 conference at which Facebook CEO Mark Zuckerberg announced their partnership. “Music is one of the most social things there is,” Ek went on, arguing that integration with Facebook would “help everyone to discover more free music than ever before.” Interestingly, piracy still loomed in the background, since Ek stated that “social discovery on Facebook means that we’re bringing people back to paying for music again.”¹⁷

Whereas chapter 1 described the historical context of this merger, what we want to accentuate here is that the social discovery of music testifies to the ways in which audio files, social media feeds, or updates are enmeshed in commercialized listening experiences. Through the widely adopted pub/sub system—a software architecture whereby publishers can transmit categorized messages to interested subscribers, without members of either group knowing the identities of or communicating directly with the others—Spotify has integrated social features to facilitate the sharing and following of music activities among users in real time.¹⁸ At the same time as the Facebook deal, Spotify also decided to open its API to external developers, whose applications could retrieve data from the Spotify music catalog. For a while, basically any third-party developer could build HTML apps through the Spotify API, although they had to “go through a rigorous Spotify approval process before being released on the platform,” as one report stated. The new potential for integration was championed as a “great achievement” but also raised some eyebrows: “Can Spotify shift its big-picture focus from sales and expansion to serving their … paying customers? No iPad app? Forced Facebook linking? … Has Spotify turned into AOL Music?”¹⁹

A number of popular music apps were developed, yet in October 2014, the so-called Spotify Apps API was discontinued. “As Spotify evolves and priorities change, we sometimes have to remove APIs from active development and shift focus to the relevant platforms,” the developer blog explained.²⁰ Data integration came with a price and was hardly frictionless. Daniel Ek was proven to be mistaken since the “frictionless sharing” and “frictionless experience” with Facebook caused a considerable public debate. Critics agreed that intentional sharing had its advantages but soon noticed that the perception of an endless list of the music being played by Facebook friends risked turning into sheer noise. “Seeing a constant flow of songs in the news ticker is at best irrelevant and at worst annoying,” as one critic put it.²¹ In September 2011, less than a week after Ek’s blog post and Zuckerberg’s keynote at the F8 conference, Spotify found it necessary to issue a clarification and instruct users on how to control their settings: “It’s entirely up to you what you decide to share and what you keep private.… To share the music you listen to with Facebook, make sure the box is checked. If you’d rather not share, just uncheck the box.” Moreover, hinting at the awkward aspects of music sharing, the blog post was later updated to announce an additional feature: “Many of our users have told us that they like to share what they’re listening to, but also want an easy way to hide their occasional guilty pleasures. So we’ve now added a new ‘Private Session’ mode to our latest update.”²²

Platformization thus had its disadvantages. The point we want to make here is that frictionless sharing with Facebook, engineered through aggregated information residing “outside” and then “pulled into” Spotify’s data infrastructure, is of utmost importance to the ways in which the service wants to be used and understood. Linked data (as well as similar concepts around the semantic web) have been a trend in a number of public and commercial domains for more than a decade. Interoperability and machine-readable data have pushed organizations, and later commercial enterprises, to try to avoid the notion of data silos and, instead, to work toward data integration and interlinked content. Commercial streaming services, such as Spotify, Apple Music, Deezer, and Pandora, are no exception. They all integrate data and metadata to create attractive services that engage audiences. Hence, when files become music at Spotify, aggregation of data occurs on, at, and via many computational layers. As we argued in our introduction, reducing Spotify to a platform makes these exchanges invisible.

Moreover, old Spotify blog entries make it apparent that such data integration has been an essential commercial strategy for Spotify, with frequent and similar alliances and mergers occurring since the Facebook deal. In April 2013, for example, Spotify and Twitter, together with iTunes and Rdio, launched Twitter #music, a service that boasted it would change the way people found music online based on Twitter activity. Tweets, likes, and retweets were supposed to “detect and surface the most popular tracks and emerging artists.”²³ At the same time, Spotify began to offer new tie-ins with the auto industry. A smaller deal with Ford came first, followed by one with Swedish car manufacturer Volvo and the launch of its Sensus Connected Touch, “a voice-activated music streaming service fully integrated with the car manufacturer’s new touch-enabled dashboard.”²⁴ Later, Spotify also collaborated with Uber, BMW (ConnectedDrive), and MINI. In March 2015, Spotify moved from cars to games, partnering with Sony and bringing music to its PlayStation console: “Listen while you play” was the slogan. And a year later, in September 2016, even love was added to the data integration list, when Spotify partnered with a well-known dating app: “Tinder, now with music by Spotify, not only gives you the capability to show your favorite artists and music preferences, but also to pick the Anthem that defines you … and maybe make a little music with someone new.”²⁵

These partnerships have been geared toward expanding the experience of using Spotify, securing the company a massive presence and visibility on different platforms and venues, and obtaining increasing amounts of both internal and external user-generated data and information. While data from listener activity and user profiles are essential for Spotify’s data infrastructure in general and its music recommendation systems in particular, some data are never shared.

Spotify claims that all information the company has generated or gained access to in recent years “simply helps us to tailor improved experiences to our users, and build new and personalized products for the future.” This quote, taken from an autumn 2015 blog post, is particularly interesting since it appears to be related to a number of new corporate alliances that had prompted (or forced) Spotify to update its “Terms and Conditions of Use and Privacy.” Brand manager Candice Katz stated that the company wanted to be as “open and transparent as possible when it comes to how we describe our business, how we work with advertisers, what information we collect, and what we do with it.” In corporate lingo, Katz assured that the privacy of “customers’ data is—and will remain—Spotify’s highest priority.” Then again, her overview of the new policy made it all too apparent what was at stake:

The Information We Collect section has been expanded to include new technical data such as additional cookies, device information, and network information.… We may ask for customer permission to collect information from new sources, such as address book, location, and sensor data from the mobile device to improve the customer experience and inform product decisions. We provide more clarity on how we share data with partners who help us with our marketing and advertising efforts for both Spotify and our brand partners.²⁶

Spotify users were upset about these sweeping privacy changes, which gave the company greater access to personal data on users’ smartphones, among other issues. The Guardian claimed that, from now on, Spotify would be able to collect sensor data, “about the speed of your movements, such as whether you are running, walking, or in transit.” Some information would be shared with advertisers, “although Spotify did not spell out exactly what data it would pass on.”²⁷ People took to Twitter and announced they were canceling their accounts: Swedish game developer Henrik Pettersson tweeted, “@Spotify account ended. I suggest you do the same. Privacy policies like that must die”; his colleague Markus “Notch” Persson (CEO of Mojang and the creator of Minecraft) retweeted this and stated, “I just cancelled mine too.”²⁸ A heated exchange ensued on Twitter, in which Daniel Ek tried to persuade Persson, his business acquaintance, to change his mind. “People are quitting Spotify over its new privacy policy,” Business Insider stated, and the Twitter conversation was picked up and quoted in a number of online media sources.²⁹

The dispute led Ek to publish a blog post famously entitled “SORRY.” He admitted that Spotify should have done “a better job” in communicating what the new privacy policies meant and how “information you choose to share will—and will not—be used.” He then described the new types of information, “including photos, mobile device location, voice controls, and your contacts” that Spotify wanted to access. But “let me be crystal clear here,” Ek asserted. “If you don’t want to share this kind of information, you don’t have to. We will ask for your express permission before accessing any of this data—and we will only use it for specific purposes that will allow you to customize your Spotify experience.” Finally, Ek also confessed that Spotify did share data with partners “who help us with marketing and advertising efforts” but assured that all information was “de-identified—your personal information is not shared with them.”³⁰

`Audio Files and Streaming Infrastructures`

We have presented a brief contextualized discussion of Spotify’s data infrastructure in order to demonstrate that it is constructed and assembled by more than just audio files. But what about the files themselves? A brief survey of data formats and streaming infrastructures here may suffice.

Spotify’s audio streaming service uses the Ogg Vorbis format, not MP3, for music. Ogg Vorbis is an open-source lossy audio compression method that offers roughly the same sound quality as MP3 at a smaller file size, which is crucial for streaming. A streamed track is always played immediately after a small amount of audio data has been received—a technological fact that our subsequent intervention will analyze in detail. Music files are not permanently stored on the destination device. For music streaming to occur, the Spotify client and the server need to communicate, with the client storing a few seconds of sound in a buffer before starting to send it to the speakers. Consequently, low latency is key to the Spotify service. (Latency is a measure of the delay between requesting a song and hearing it; “low latency” means that when a user presses play, a track should start more or less instantly.) To achieve this, the Spotify web client fetches the first part of a song from its infrastructural back end and starts playing a track as soon as sufficient data has been buffered as to make stutter unlikely to occur. Therefore, “the main metric of the Spotify storage system is the fraction of requests that can be served with latency at most t for some small value of t, typically around 50 [milliseconds].”³¹ There are different quality ratings for streaming: normal mobile quality (96 kilobits per second, or kbps), desktop and web player standard quality (160 kbps), and high quality (320 kbps), which is available only to Premium subscribers. It is also possible to make Spotify one’s own “all-in-one music player.” In addition to the more than thirty million tracks that Spotify offers from its catalog, “you can also use your Spotify app to play music files stored on your computer,” as MP3, MP4, and M4P audio formats are also supported.³²

As we have shown in chapter 1, the technical roots of the service are within file sharing and music piracy. Sharing files via Napster or the Pirate Bay, however, was never only about the actual exchange but always involved a networked sociability with actors and/or people. In a similar sense, music files on Spotify engage in a form of data sociality, resulting in assemblages of different digital assets (such as metadata) that usually include multiple formats. Spotify’s tech stack includes hundreds of programs and services, from Amazon and Apache to Python and WordPress, engaging with one another. If the Spotify back-end architecture is heavily service-oriented (mostly written in Python or Java), audio files are also retrieved from remote and different locations, with streaming coded as a combination of server-based access to audio files (on Amazon S3) and peer-to-peer (P2P) technologies (prior to 2014). Spotify used to have at least four data centers around the world, including one in Stockholm and one in Ashburn, Virginia. The first characters of a server’s hostname, such as “ash2-dnsresolver-a1337.ash2.spotify.net,” indicate the data center’s physical location (in this case, Ashburn).

In order to facilitate and sustain its infrastructure, Spotify ran one of the largest P2P networks on the internet for a number of years. According to Spotify, during that period, less than 10 percent of music playback came from its own servers, with roughly 35 percent coming from P2P networks and an astonishing 55 percent from the user’s local cache. The exception was playout from smartphones, where all music was streamed directly from Spotify servers. At a presentation in 2011, Ricardo Vice Santos, then head of growth/new markets at Spotify, described streaming music in the following way: “Request first piece from Spotify servers | Meanwhile, search for peers with track | Download data in-order | When buffers are sufficient, switch to P2P | Towards end of a track, prefetch next one.”³³ Spotify needed P2P to guarantee that all tracks could be played with the lowest latency possible. With a dedicated P2P system, a streamed song, played for the first time, would therefore be stored on a user’s hard drive (cache). When playing the track again, the cached version was used instead of repeated downloads from the network, and the client would (via P2P) be able to use parts of a song from other users. When a client played a music track, data was thus obtained from a combination of three sources: “the client local cache (if the same track has been played recently), other Spotify clients through peer-to-peer technology, or the Spotify storage system in a backend site.”³⁴ All of this happened in the background, however, and contributed to a smooth user experience.

Only in the spring of 2014 was it announced that Spotify would shut down its P2P servers. “We’re now at a stage where we can power music delivery through our growing number of servers and ensure our users continue to receive a best-in-class service,” the company stated.³⁵ In order to upgrade its infrastructure, Spotify started to buy or lease data center space, server hardware, and networking gear. In 2016, the infrastructural framework changed again. After a few years of operating its own data centers and running its “core infrastructure on [its] own private fleet of physical servers (aka machines) rather than leveraging a public cloud such as Amazon Web Services,” it now announced it would start working with the Google Cloud Platform team to provide a new streaming infrastructure.³⁶ Spotify now determined storage, computing, and network services available from cloud providers to be sufficiently robust (and low cost). As Spotify declared, “Good infrastructure isn’t just about keeping things up and running, it’s about making all of our teams more efficient and more effective, and Google’s data stack does that for us in spades.”³⁷

`Aggregating Content`

The brief survey we have provided demonstrates that turning files into music is a process that involves an exceedingly interrelated data stack and a complex streaming infrastructure of software services, metadata, and user-generated data. Spotify’s data infrastructure is built on layers of interrelated services, streams, and exchanges. While postings on Facebook and elsewhere are important for Spotify’s claims about making music more social—and while metadata flows (via the Echo Nest) are pivotal for blurring the boundaries between music and information about music—aggregation of actual content forms yet another important data layer of Spotify’s online presence. In “Intervention: Record Label Setup,” we described how our self-produced music helped us gain a number of insights into music aggregation procedures. In what follows, we will complement these insights with some considerations of how music aggregation and databases containing cultural content are conceptually structured in ways that enable certain forms of use and experience while disabling others. Spotify depends on aggregation practices. When files become music on Spotify, the aggregation of content is the first step.

Aggregation is a generic term for the internet’s capacity to pull content from various sources and make it accessible at dedicated sites, such as Netflix, Wikipedia, and Spotify. While aggregation practices existed long before the internet—making cultural content available for free or a small fee is what libraries, circuses, and museums have done for centuries—online aggregation is often related to the internet’s propensity to scale. Pulling together information into one single location is increasingly important online, and when it comes to music, it is aggregation—rather than the process of copying bits, called streaming—that shapes interfaces and listening experiences. Aggregation is not the source of the stream but a facilitating principle that unites the distinct data particles into a coherent whole.³⁸

Music aggregators act as intermediaries that connect smaller rights owners to streaming services, while major conglomerates and labels usually entertain a direct business relationship with the service. In the case of Spotify, this is unsurprising, since the company is partly owned by Universal, EMI, and Warner. As intermediaries, aggregators intrude into, and fundamentally alter, existing distribution chains or information channels. According to Daniel Johansson, music aggregators can therefore be perceived as a new function within the music industrial system. Johansson estimated in 2013 that about forty music aggregators were operating in the market—a figure that may have risen substantially since then.³⁹ In many ways, however, today’s digital distribution chain—label, aggregator, retailer—is reminiscent of the way in which recorded music on physical carriers used to be distributed: from a label to a record store via a distributor.

As a facilitating principle, aggregation stands for a techno-social trend that combines an ever-expanding scope of culture, knowledge, and commodities with completely new systems to sort, aggregate, and filter content. Aggregation is an example of what Jeremy Wade Morris calls infomediaries, “organizational entities that monitor, collect, process and repackage cultural and technical usage data into an informational infrastructure that shapes the presentation and representation of cultural goods.” In Morris’s view, the emergence of infomediaries marks a shift toward database and algorithmic technologies influencing not only “the organization of digital goods” but, more importantly, the curation of culture through “mining, repackaging and provision of data.”⁴⁰

Elsewhere, we have argued that as a business operation, music aggregation connects traditional institutional ideas of accumulation—think of the library and catalog metaphors used to explain online services—to the promise of new media.⁴¹ What characterizes aggregation, however, is not so much the effects of aggregation per se but the effects of the difference between various aggregation practices, not least financially. Pricing mechanisms among music aggregators, for instance, vary substantially. There are also significant differences in pricing policies between aggregators working with record labels and those that target individual artists. Some charge a small amount per played track, others an initially fixed price or a fixed price per song or album upload, and still others a certain percentage of royalties.

In its current use, the term aggregation covers a broad spectrum of industry practices and technologies, yet exactly what the term denotes remains blurry and hard to delimit vis-à-vis distribution, programming, or syndication. As Patryk Galuszka has argued, aggregators “exist in several industries, but in the digital music market their role is largely unnoticed.”⁴² Few have heard of Awal, for instance, an aggregation service promoted by Spotify that promises to get one’s music distributed “without upfront fees to over 200 different digital stores and services around the world.” With offices in Atlanta, Berlin, London, Los Angeles, Miami, and Nashville, Awal in many ways lurks in the background of the contemporary music industry, offering a “simple digital distribution license” while charging 15 percent of revenues generated.⁴³

An aggregator such as Awal bundles digital and intellectual property rights—both copyrights to sound recordings and artists’/performers’ rights—and delivers content to digital music stores, either in the form of downloading (iTunes) or streaming (Spotify). Aggregators hence operate “on the business-to-business market, where one group of contractors are record labels or individual artists, and the other group are digital music stores.”⁴⁴

If one takes the community conversation about Spotify as an indicator, it usually takes about two weeks for a new tune to go through the aggregation process and reach the user. In order to ease and inform what aggregation means, Spotify has created an artist’s guide “to walk you through setting up your artist profile, customizing your presence, and growing your fanbase on Spotify.” The notion of “delivery” is the first category one encounters in this “Spotify for Artists” guide, which enthusiastically asserts that “getting your music on Spotify is easy.” If an artist is already signed to a label or has started using an aggregation service, “they’ll get your music on Spotify for you. If you don’t, we have deals in place with a number of companies who can deliver your music to us and collect royalties for you.” At the time of writing, Spotify lists six aggregation partners: TuneCore, CD Baby, EmuBands, Record Union, Spinnup, and Awal. They are described as aggregator services that handle “the licensing and distribution of your music,” and administer royalties and monetary compensation “when your fans stream your music on Spotify.” Spotify also stresses that these services do not work for free; usually there is “a small fee or percentage cut involved … so be sure to do a little homework before picking one.”⁴⁵

Looking at the offers from music aggregators, it becomes clear that, from their perspective, Spotify is one of the dominant players on the streaming market. Together with iTunes, Amazon, Apple Music, and Google Play, Spotify is heavily promoted by aggregators such as TuneCore, CD Baby, EmuBands, MondoTunes, and RouteNote. Spotify is also integrated with delivery platforms such as FUGA, Consolidated Independent, and the Merlin Network—all of which are also promoted on the “Spotify for Artists” site: “Deliveries via these platforms will speed up the time it takes to get your content on Spotify.”⁴⁶

One of these aggregators, TuneCore, boasts a “strong partnership with Spotify.” TuneCore offers “Daily Spotify Trend Reports” as a way to keep track of “your success”—or failure, as in the case of our own recordings—and thus to be able to make timely business decisions. In addition, the “Spotify Verified Artist Accounts” allow TuneCore artists and labels “to easily build a community of followers and interact with fans directly from artist pages.”⁴⁷ The TuneCore FAQ also explicitly tries to answer the tricky question of how much money can potentially be made through Spotify sales and its “paid streams.” Spotify has deals with rights holders in all of the countries where the service is available, TuneCore asserts. “A royalty is based on how frequently your music gets played. Each stream earns you a share of Spotify’s advertising revenue.” The actual amount varies substantially and depends on “the ratio of advertising revenue and your percentage of the total number of streams on Spotify in a given month.”⁴⁸

For all types of streaming music, it should be stressed that aggregation in principle means that the distribution chain gets longer for smaller repertoire owners, as compared to major rights holders. It reduces revenues flowing back from streaming services, and in many cases, individual (and less popular) artists will find that making their tracks available through pricey aggregation services causes them to operate at a loss (since revenue per played track is almost insignificant). After all, Spotify monetizes usage, not units. On the Spotify Community Blog, there are consequently a number of complaints regarding aggregation and revenue, ranging from splitting royalties to aggregators refusing to remove music.⁴⁹ A Google search for “best music aggregator” will bring up a number of postings with similar pros and cons. Consider, for example, the following comments on a Reddit thread:

I’ve been using CDBaby for over 10 years. It was the only option for a long time. One-time setup fee and they also handle rights management if you do the pro package. HUGE time saver.

I use CDbaby too. A little more expensive to get started with, but I appreciate that I literally pay once and stuff is up forever.

I think these days most services are good, the only thing that could tilt your choice is how you release your music. If you put out a new single every week CDBaby might get very expensive for you and maybe another one of the flat fee services would be better.

tunecore is easily the best. most stores, most transparent, most information on their site, most high profile artists, etc.

But tunecore makes you pay for every upload, which some people don’t like. I for one don’t make a lot per tracks, so I rather they took a percentage of my revenue like amadea music does.⁵⁰

Music aggregators are thus a new group of intermediaries that have played a vital role in articulating the value of music over the past decade. Forming part of a wider techno-social ensemble that constitutes today’s digital media infrastructures, music aggregators have also contributed to an experience of devaluation that these services were designed to fight. “The smaller and lighter the universal music library becomes, the heavier it seems to pull us down,” as Geert Lovink has stated.⁵¹

Aggregation of sounds from a number of online distributors scattered across the internet again stresses that Spotify’s data infrastructure needs to be grasped in its heterogeneity and in its capacity to scale. Streaming services would not have been able to gain rapid, global popularity if it had not been for their aggregated, abundant back catalogs of content. In countless interviews, Daniel Ek has emphasized the importance of building a vast music catalog: “With music, rediscovery is a critical part of how you listen to music.”⁵² Consequently, during the past decade, the number of tracks available from music retailers and streaming sites has been promoted to entice prospective listeners. In many ways, Apple’s iTunes Store set the tone by persistently advocating its swelling back catalog. “The iTunes Music Store in the US, UK, France and Germany offers an extensive music library of over 700,000 songs in each country,” Apple boasted in 2004.⁵³ Today, all streaming services on the market make similar claims: Spotify and Apple Music now have well over 30 million songs, Pandora claims to feature 40 million tracks, and SoundCloud gives access to no less than 120 million user-added tracks. More music, indeed, seems to be better music.

Every stream means potentially increased revenue from advertisers in Spotify’s ad-supported version. Since Spotify Free operates similarly to commercial radio, more streams are equivalent to more usage, which is what attracts advertisers—an issue to which we will return in chapter 4. We have argued elsewhere that this is one of the reasons why streaming services such as Spotify aggregate almost indiscriminately and are more likely to include, rather than reject, various forms of (semi-)automated music and sounds—especially compared to retailers such as iTunes, from which users purchase individual downloads.⁵⁴ One might be surprised to find that Spotify contains tracks with titles such as “Aircraft Lavatory Ambience,” “Weight Loss Hypnosis,” “Car Alarm on City Side Street,” “Beach Rain,” and “Spend Less-Stop Wasting Money Subliminal Message Therapy”—not to mention the one-hundred-track album Correct Wrong Sound Effects. While these tracks do not attract crowds of listeners, they are an important part of the marketing hype around “more music” and thus part of the “all-you-can-eat” bid that streaming services offer. The multifaceted and unregulated market of music aggregation is the main reason why all these tracks are available.

As we have previously described in the intervention on our record label experiments, current rejection criteria at music aggregators are more or less arbitrary, depending on whether users pay a fee or not. Aggregators disagree about whether the same tracks and albums even count as music. The line between music and nonmusic, artist and machine becomes blurred. Bizarre tracks on Spotify—such as “Overcoming Job Loss – Positive Affirmations”—might not come across as music, but they have passed an aggregator. Still, record labels and artists expect Spotify to act as a walled-off streaming service with professional offerings rather than a semi-open platform. Consequently, streaming fraud has generated worries within the music industry, not least since music hacks, pranks, and tricks tend to get public attention. These have included the funk band Vulfpeck and its conceptual album, Sleepify (which contains five minutes and sixteen seconds of pure silence); the band Ohm & Sport and it application Eternify, where—for a very short time—one could enter the name of a favorite artist and play songs on repeat in thirty-one-second intervals in order to maximize the artist’s revenue; and the music spammer Matt Farley, who has personally released over fifteen thousand songs.

The spread of illicit or non-“organic” promotion in the form of automated listeners on Spotify and other services may be seen as relating to similar aggregation devaluation mechanisms—and to the paradoxically expanding back catalog of unheard music. Approximately 20 percent of Spotify’s catalog has not been listened to by anyone even once. Thus, more music always means more unheard music. A site such as Forgotify testifies to the vast amount of aggregated but undiscovered content. “Millions of songs on Spotify have been forgotten. Let’s give them new life in new ears—yours. [At Forgotify] we were so shocked to learn that millions of Spotify songs had been played only partially or never at all. A musical travesty, really. So we set out to give these neglected songs another way to reach your earholes”⁵⁵ The dark side of aggregation equals plenty of “zombie music” on Spotify.

`The Radio Loop Experiment`

When music becomes data and comes to resemble all other kinds of digital content, it also comes to adhere to all aspects of computational logics, even the more annoying ones. Therefore, we would like to conclude this chapter by describing one experiment—related to the previous discussion on aggregation—that we conducted at Humlab. The experiment proposes a set of methodologies for performing humanist inquiries on “midsize” data and black-boxed media services, such as Spotify, that increasingly serve as key delivery mechanisms for cultural goods. As with some of our prior interventions, the experiment used bots as research informants and sought to critically investigate aggregation of music and the prospective lure of infinite archives via the radio functionality of Spotify.

Through unknown algorithms, Spotify Radio offers users a potentially unlimited avenue of music discovery due to the service’s vast back catalog of aggregated content. The service, however, has been severely criticized: “Is Spotify Radio broken?”; “How do I get Spotify to stop playing the same few songs for every artists [sic]?”; “How do I teach a Spotify radio station to play a wider array of songs?”; “Is the Spotify streaming radio … purposefully terrible with the intention of trying to get people to upgrade?”⁵⁶ As these queries from the website Quora indicate, not only do many users dislike Spotify Radio, they have even accused it of playing the same artists over and over.

We have also discussed similar issues with the poor performance of the Spotify Radio algorithm. Such assumptions reveal a normative claim that the radio algorithm should produce apt recommendations. In order to answer at least some of these issues, we decided to set up an experiment that would explore Spotify Radio. Our objective was to uncover why we rarely liked the songs that the radio algorithms suggested we should like. But given normative assumptions about the ways in which Spotify Radio ought to work, the research question also hinted at the ways in which algorithmic music discovery today features and promotes some artists—and obstructs others. Hence, a software-driven cultural analysis of music delivery mechanisms could potentially reveal the algorithmic flaws that regulate music recommendations to the detriment of a more diverse listening experience, making less room for emerging musicians or neglected genres (with economic ramifications).

Our experiment attempted to analyze the possible limitations found within “infinite archives” of music streaming services such as Spotify. For commercial reasons, Spotify Radio claims to be both personalized and never ending. Our hypothesis, however, was that Spotify Radio did not consist of an infinite series of songs. Rather, music seemed to be delivered in limited loop patterns. If our hypothesis held true, what would such loop patterns look like? In order to answer these questions, we set up an experiment with the purpose of examining Spotify Radio loops. Our loops were constructed using 160 “bot listeners.” All of our bots were Spotify Free users with literally no track record—they had “heard” no music before they were put into action. Our primarily interest was not the personalized recommendations that Spotify’s algorithms offered but rather how Spotify Radio functioned generically. Moreover, providing all of our bots with a personal track record would have been impractical, if not impossible, to accomplish. In addition, as virtual informants, our bots did not explicitly collect information. They were programmed and designed to search for a track, retrieve subsequent songs, partially interact, and (most importantly) log data caused by different “actions.”

If one of our aims with the experiment was to study the repetitiveness in loop patterns, another hypothesis was that the size and structure of radio loops might depend on music genres as well as popularity. We decided to let our bots “listen” to both a hit song and a less popular track with some contextual similarity. Then, all of our bots would start a radio channel based on Swedish music from the 1970s and listen for twenty-four hours. A major and a minor experiment were conducted. In the first round, 120 of our bots (although many eventually failed because of various technological problems) started Spotify Radio based on the highly popular ABBA song “Dancing Queen” (1976), which has been streamed some sixty-five million times on Spotify. The second round of bots (40 in all, with a few failures) started a radio channel using the significantly less popular “Queen of Darkness” by Swedish progressive rock band Råg i Ryggen (1975), with approximately ten thousand streams. The bots were to document all subsequent tracks played in the radio loop, as well as to interact differently within the Spotify web client as an “obedient” bot listener, a “liker,” a “disliker,” or a “skipper.” The interactions were documented, including tracks and artists played, as well as breaks for advertisements.

Elsewhere, we have described in detail the results of our Spotify Radio experiment.⁵⁷ The first thing to note is that it is possible to measure loop patterns on Spotify Radio. Working with bots as research informants allowed us to monitor the logs they produced and thereby empirically sustain claims of repetitiveness within Spotify Radio. The regularity of patterns was clear, and music loops were definitively not endless. On the contrary, they displayed a repeated pattern with only slight variations according to which artist a radio station was based on. For instance, the specific track that we based our radio stations on kept returning in the bot playlists. If a radio loop started with “Dancing Queen,” it was played again by the Spotify Radio algorithms after about fifty tracks. Bots listening to a radio station based on “Queen of Darkness” displayed a similar tendency, but with the difference that the song was not repeated as often as “Dancing Queen” and at longer intervals (regularly after some seventy tracks or so).

A preliminary conclusion to draw from the experiments with Spotify Radio is that similar artists reappeared frequently within all bot playlists. Music recommendation algorithms at Spotify did not really take advantage of the archival infinity of the service. Thus, if Spotify Radio is about personalization of content, as the company claims, then the recommendation algorithms were a disappointment. An even more troubling result, at least for Spotify, was that radio loops tended to look more or less the same, independent of bot characteristics. Giving Spotify Radio the user feedback of “thumbs up” (like), “thumbs down” (dislike), or skip did not produce significant differences in the results.

One result of the experiments was therefore that music loop patterns basically looked the same, regardless of interaction within the Spotify web client. Another general conclusion was that the recommendation ability of Spotify Radio is exaggerated. The claim that “the more you personalize, the better the music gets” should be perceived as a mendacious company claim used to attract listeners and stir commercial interest in the radio functionality (when the service was new). Yet, since complaints were made right after the launch of Spotify Radio, it is likely that the recommendation functionality was flawed from the start. The public critiques on Quora of the inadequate functionality of Spotify Radio were thus spot on.

Today, the technology seems to have caught up—but not in a radio setting. In recent years, Spotify has put way more emphasis on its “Discover Weekly” and “Release Radar” playlists than on its radio functionality. In short, there seems to be a particular tech-musical recommendation narrative stretching from Spotify Radio (2011) to “Discover Weekly” (2015) and “Release Radar” (2016). The latter is different from “Discover Weekly” in that its tracks are brand new and have no listening data. Instead, Spotify relies on a solution that tries to predict who will enjoy a song by analyzing the audio signal. In the end, it seems that traditional radio recommendations appear to be less significant for Spotify, at least in comparison to other, newer types of suggestions. As streaming music—and not radio—has become the default listening mode, it is hardly surprising that the radio metaphor would gradually lose its popularity and consequently be largely replaced with new computational recommendation formats based on taste profiles, song identification, and digital fingerprints. Music, in other words, is now made of other kinds of files.

`Notes`

5. This type of scholarly literature on computational media is naturally heterogeneous and vast—but also driven by a common interest, even if separate media modalities are investigated. Publications that have inspired us include the following: Matthew G. Kirschenbaum, Mechanisms: New Media and the Forensic Imagination (Cambridge, MA: MIT Press, 2008); Wendy Hui Kyong Chun, Programmed Visions: Software and Memory (Cambridge, MA: MIT Press, 2011); Jonathan Sterne, MP3: The Meaning of a Format (Durham, NC: Duke University Press, 2012); Wolfgang Ernst, Digital Media and the Archive, ed. Jussi Parikka (Minneapolis: University of Minnesota Press, 2013); Richard Rogers, Digital Methods (Cambridge, MA: MIT Press, 2013); and Robert W. Gehl, Reverse Engineering Social Media: Software, Culture, and Political Economy in New Media Capitalism (Philadelphia: Temple University Press, 2014).