Conference Agenda

AI music platforms such as Suno and Udio allow users to generate new music by typing in text prompts describing genre, style, and mood, rather than by notating a score or recording an instrument or voice. Developing these generative models involves an iterative training process between humans and machines (Seaver, 2022). Training a generative model usually starts with a dataset that includes audio files paired with text annotations, establishing the “ground truth” of these classifications. By correlating these two elements, the model “learns” what the audio is. Rather than directly listening to the audio file and hearing it as a collection of sounds that can be described by music theory or in terms of specific musical practices, AI models “hear” the audio dataset as a collection of words — provided by humans who previously listened to and labeled the music.

This paper explores the history of music metadata and how that history shapes metadata’s role as a new epistemology of generative music. From nineteenth-century practices of classifying and publishing music for melodrama and early-twentieth century silent film cue sheets, to current music licensing marketplaces and libraries, so-called “functional music” pairs music with narrative storytelling. This functional music, I argue, is the foundation of our generative music systems. Prompting a generative system to create a “sad song” may seem to rely on internal musical structure. Yet I show how these systems rely on much older encodings of genre, style, and mood from the late-nineteenth and early-twentieth centuries. The assumptions and tropes embedded in and arising from these practices are now the ground truth of generative AI music systems. Drawing on industry experience, ethnographic work at several music technology companies, and close readings of silent film cue sheets and musical datasets, this paper argues that these much-hyped contemporary systems can best be understood as an evolution and continuation of functional music.

The hip-hop of artists AgusFortnite2008, Acid Souljah, and xaviersobased is hard to classify. The musical surface is saturated to the point of incomprehensibility; an unrelenting barrage of audio internet memes, foley effects, and distorted 808 drum samples are superimposed over the rap vocals. Critical response to this music has tended either to dismiss it as nonsense, or else to justify its nonsense through assertion of high art parallels like Dadaism. In a 2024 article, critic Kieran Press-Reynolds labels this genre with the portmanteau “Shitpost Modernism,” drawing a correspondence between the “high-low anarchy” of reference mixing shared between the musical genre, the internet practice of “shitposting,” and traditional accounts of postmodernism.

Both dismissive and celebratory viewpoints, I argue, obscure the actual aesthetic strategies at play in this repertoire, however. Instead, as I demonstrate in this paper, the distinguishing feature of this music is its importation of aesthetic values from internet meme practices–values that are irreducible to previous understandings of postmodern pastiche, but that also subvert aesthetic priorities of much prior and contemporary hip-hop. The overstimulating meme montage is a longstanding internet practice, from the early YouTube trends YouTube Poop and Montage Parodies to the more recent TikTok phenomenon “21st Century Humor.” These texts derive humor from remixing and recombining more straightforward internet memes into a mind-boggling surfeit of reference points. Where the unlikely image pairings of Dada photomontage encouraged the onlooker to draw allegorical and political connections, the meme montage gleefully overwhelms the viewer's interpretive ability. Unsurprisingly, the translation of this meme technique into hip-hop produces unruly results. The aesthetic priorities of hip-hop are typically given to be flow, groove, and lyricism. Through incorporating the practices of internet meme montage, these artists disrupt their received aesthetic mandates and forge a new style of disorientation and irreverence. Drawing on music theoretical studies of meme practice (Boone 2023, Shelley 2020), theorizations of play and surface through topics and interfaces (Moseley 2015, Allanbrook 2014), and recent work in cultural criticism (Kornbluh 2024, Ngai 2020), this paper analyzes “Shitpost Modernism” with an eye toward formal play, genre formation, and the unmaking of musical meaning.

When theorists think of the “musical surface” they likely think of topic theory – that “site where musical mimesis plays itself out” described by Wye J. Allanbrook, set in opposition to Romantic notions of musical depth (Watkins, 2011). Yet the musical surface appears in another, unexpected set of literature: early-twenty-first century Music Information Retrieval (MIR). In a statement characteristic of the time, George Tzanetakis et al. (2001) defined the musical surface as “the characteristics of music terms related to texture, timbre, and instruments,” calculable through “the statistics of the spectral distribution over time.” For many in MIR, this surface was sufficient for querying audio files. They assumed that identifiable musical characteristics could become known through statistical extractions from audio files — the machine-audible surface of the sound. While these two formulations seem to have run in parallel, this paper shows that they share a convergent intellectual history, one reaching back to the 1950s cybernetic turn, when cybernetics became the “universal discipline” (Bowker, 1993).

This paper argues that digital systems have transformed the musical object into something new. This transformation, however, has been occluded. Tracing this history enables a better understanding of the “being” digital systems have made of musical objects. I show how Leonard Meyer and Leonard Ratner, immersed in inescapable cybernetic discourse, both developed anti-Romantic, computational views of music. Moreover, I trace an unlikely complex of interchanges between their students, including Robert Gjerdingen and Kofi Agawu, and prominent MIR researchers, like George Tzanetakis. MIR researchers found the views of contemporary music theorists of eighteenth-century music conveniently compatible with their own: music as modular, syntactical, and knowable through an audible surface. Moreover, computational models learned a particular kind of musical surface — one derived almost exclusively from collections of Euro-American music. In the process, Euro-American music became the standard against which all other music would be compared. Understanding the “digital musical object,” I argue, requires us to revive older, outmoded theories of eighteenth-century music. Many Romantic legacies persist in our ideas about music. Yet our digital systems compute something much closer to an eighteenth-century understanding: music as meaningful sonic surface.