Artificial Intelligence & Popular Music: SKYGGE, Flow Machines, and the Audio Uncanny Valley

Abstract

This article presents an overview of the first AI-human collaborated album, Hello World, by SKYGGE, which utilizes Sony’s Flow Machines technologies. This case study is situated within a review of current and emerging uses of AI in popular music production, and connects those uses with myths and fears that have circulated in discourses concerning the use of AI in general, and how these fears connect to the idea of an audio uncanny valley. By proposing the concept of an audio uncanny valley in relation to AIPM (artificial intelligence popular music), this article offers a lens through which to examine the more novel and unusual melodies and harmonization made possible through AI music generation, and questions how this content relates to wider speculations about posthumanism, sincerity, and authenticity in both popular music, and broader assumptions of anthropocentric creativity. In its documentation of the emergence of a new era of popular music, the AI era, this article surveys: (1) The current landscape of artificial intelligence popular music focusing on the use of Markov models for generative purposes; (2) posthumanist creativity and the potential for an audio uncanny valley; and (3) issues of perceived authenticity in the technologically mediated “voice”.

It is often said that television has altered our world. In the same way, people speak of a new world, a new society, a new phase of history, being created—‘brought about’—by this or that new technology: the steam engine, the automobile, the atomic bomb. Most of us know what is generally implied when such things are said. But this may be the central difficulty: that we have gotten so used to statements of this general kind, in our most ordinary discussions, that we can fail to realize their specific meanings.

—Raymond Williams (1974)

To echo the words of Raymond Williams, and to begin so boldly: We are on the edge of a new era of popular music production, and by extension, a new era of music consumption made possible by artificial intelligence. With the proliferation of AI and its developing practices in computational creativity, the debates surrounding authorial meaning are increasingly convoluted, and the longstanding tradition of defining creativity as an innately human practice is challenged by new technologies and ways of being with both technology and creative expression. While the questions over intellectual ownership, authorship, and anthropocentric notions of creativity run throughout the myriad forms of creative expression that have been impacted by artificial intelligence, in this article I explore, more specifically, the new human-computer collaborations in popular music made possible by recent developments in the use of Markov models and other machine learning strategies for music production, and use this overview as a basis for introducing the idea of an audio uncanny through a case study of SKYGGE’s Hello World.

There is a much longer history of artificial intelligence music in the art music tradition (Miranda 2000; Schedl et al. 2016; Roads 1980), but its use in popular music is currently predominantly that of novelty, experimentation, and largely as a tool for collaboration. The purpose of this article, therefore, is to document the present practice and understanding of the use of artificial intelligence in popular music, focusing on the first human-AI collaborated album, Hello World, by SKYGGE, utilizing the Flow Machines technology and led by François Pachet and Benoît Carré. Using SKYGGE as a case study creates space to discuss some of the current perceptions, fears, and benefits of AI pop music, and builds on the theoretical discussions surrounding AIM (artificial intelligence music) that have occurred with the more Avant Garde uses of these technologies. As such, in this article I survey (1) the current landscape of artificial intelligence popular music focusing on the use of Markov models for generative purposes; (2) posthumanist creativity and the potential for an audio uncanny valley; and (3) issues of perceived authenticity in the technologically mediated “voice.” These themes are explored through a case study of two tracks from Hello World: “In the House of Poetry” featuring Kyrie Kristmanson and “Magic Man.”

Considering the relative newness of AIPM (artificial intelligence pop music), there is limited research on the subject, which this article addresses. By building on work from the AIM tradition, AIPM marks a separation of not only technique and audience reception/engagement, but also concept. Whereas AIM has largely been concerned with pushing the boundaries of possibility in composition, AIPM is currently being used predominantly to speed up the production process, challenge expectations of creative expression, and will predictably mark a key shift in music production eras: analog, electric, digital, and AI. In many ways, these differences run parallel to other distinctions that arguably still exist between the art and popular worlds: academic/mainstream, art/commerce. The use of these technologies creates shared practice in the surrounding discourse, but their use also marks differences in approaches to the creative process.

AIPM has not had its “breakthrough” moment yet, and it’s likely that it never occurs in such a momentous fashion as Auto-Tune’s re-branding as an instrument of voice modulation in Cher’s “Believe” (1996), subsequent shift into hip hop via T-Pain, and “legitimization” through other artists such as Kanye West and Bon Iver (Reynolds 2018; Browning 2014). These key moments in popular music history have become mythologized, marking the rise of Auto-Tune as not only a shift in production method, but also our relationship to, and understanding of, the human voice as authorial expression, sparking some of the first debates about posthumanistic music production and cyborg theory in musicology (Omry 2016; Auner 2003; James 2008).

In 2019, Auto-Tune is mundan.1 Following the cycle of most new technologies—the synthesizer, digital audio workstation (DAW), electronic and digital drum machines—Auto-Tune had its moment of disruption, its upheaval period before settling into normative practice (McLuhan 1966). It is yet to be seen how AIPM also navigates this process and how, like those that came before it, the technology evolves to incorporate unintended uses within the music industry. It is those moments of unintentionality that are often the most captivating and push the industry forward, not unlike Auto-Tune and its now ubiquitous presence as a vocal manipulator, but of course, the most well-cited case; the birth of hip hop through unintended uses of turntables. It is often these initial moments of novelty that spurn new genres, practices, and trajectories in popular music. It is important to document these initial moments of AIPM, before it inevitably reaches the next stage of development, and extends into those unanticipated uses.

1. Situating Artificial Intelligence Popular Music

The history of AIPM extends from the history of AIM, and overviews of that history have been addressed elsewhere (Miranda and Biles 2007). While SKYGGE’s Hello World is documented as the first full album to be a true collaboration between AI and human production techniques (Avdeeff 2018), there are prior instances in popular music history that anticipate this album, such as David Bowie’s use of the Verbasizer on Outside (1995). Although not a tool for music production, the Verbasizer is a text randomizer program that automates the literary cut-up technique, in order to alter the meaning of pre-written text through random juxtapositions of materials (Braga 2016). In addition, since 2014, Logic has had the option to auto-populate drum tracks for users, and more recently they have improved upon this feature by adding quantization that mimics a more “human” sense of timing based on minute variations vs. strict adherences to tempo. Furthermore, and most notably, Amper, the “world’s first AI music composer,” released its beta software in 2014 and functions through human and AI collaboration to create new music based on mood and style. At the time of writing, the beta platform has been taken offline and the enterprise version, Amper Score, is soon to be released. However, the main market for Amper Score are companies/individuals who require royalty-free music to accompany branded content, such as podcasts, promotional videos, and other content that would normally use copyrighted stock music. Similar capabilities can be seen in Jukedeck, also founded in 2012. Current uses of Amper and Jukedeck are to speed up the process of creating royalty-free stock music, or even Muzak, and they are yet to make an impact on the production of mainstream popular music. They are generally used to produce music that is meant to accompany other audio-visual content, and therefore, does not need to be highly emotionally engaging, or even, arguably, incredibly musically interesting. Regardless of the potential cost-, time-, and energy-saving impacts associated with AI-aided music production, if the added benefit of using AI to produce background content ultimately results in the increased quality of said music, I see no drawbacks.

Schedl, Yang, and Herrera-Boyer define the canon of music systems and applications that are utilized to automate certain music production or music selection processes for online stores and streaming services as intelligent technologies, while also acknowledging the problems in defining something as intelligent (Schedl et al. 2016). As there is no definitive definition for what intelligence entails, the word is used more often as a marketing technique than a description of what a product/software/platform can do, functionally speaking. As they note, “musical intelligence could either be a capability to build interconnections among the different layers of the music work…to generate sequences that other systems consider as ‘music,’ or to anticipate musical features or events while listening to a musical performance, or all of them at the same time” (Schedl et al. 2016). They therefore build a typology of examples of intelligent music systems, which includes: systems for automatic music composition and creation, such as CHORAL, Iamus, Coninuator, OMax; systems to predict music listening, such as Just-For-Me, and Mobile Music Genius; systems for music discovery, such as Musicream; and algorithms to curate music based on mood/emotions. Further tools used in current AI assisted popular music production include: IBM Watson, Magenta, NSynth, AIVA, and Sony Flow Machines.

AI music production has taken a variety of forms, from algorithmic to the use of neural networks and machine learning. As the basis for Flow Machines is proprietary generative constraints-based Markov methods, such is the focus of this article. Fundamental to all methods, though, is the use of statistical analysis, or a set of rules and probabilities, which are then used to predict and create new materials based on what has been created before. Algorithmic music, which has a much longer history and use in Western art music, is less relevant to a discussion of AIPM, but it should be noted that there has long been an intimate connection between music and algorithms, as many musical formats are also bound to particular sets of rules, or algorithms. Rule-based compositions, such as theme and variations, or the 12-bar blues, can be defined as algorithmic music, because of their dependence on specific rules as the basis for composition, although aleatoric, or chance music, has more commonly been associated with algorithmic production.

Markov chains and other forms of machine learning composition rely on a corpus of existing music, scores, and/or audio, that are used as the basis for new compositions. Algorithmic compositions differ from machine learning composition in that algorithmic music accomplishes a predefined goal, whereas generative modelling is not necessarily accomplished by means of a pre-determined algorithm. These terms necessarily overlap, but differ in their aims. For example, algorithmic music is music generated by defined algorithms. Markov chains produce randomly distributed musical elements through stochastic models based on an analysis of a defined corpus, and can rely on a much smaller dataset than machine learning processes. According to Charles Ames, a Markov chain “models the behavior of a sequence of events, each of which can assume any one of a fixed range of states. Changes in state between consecutive events commonly are referred to as transitions” (author’s emphasis) (Ames 1989). It should be noted that, although there is an act of “creation” made possible through AI software, that creation is nevertheless dependent on the music of human origin, which is used as the basis for analysis. Google DeepMind’s explorations of iterations on AI iterations for visual content hint at a possibility of entirely AI-driven audiovisual content. It will be interesting to see how these processes may be taken up in music production, as well. I question whether iterations on an AI-collaborated musical corpus will fail to engage human tastes in meaningful ways, and/or push human creativity and listening practices into new directions.

Amongst others, Curtis Roads traces the use of Markov chains for generative modelling of music to early instances in the 1950s–1960s (Roads 1985), most notably, Hiller’s fourth movement of the Illiac Suite (1956). Many of these early uses of computers in music composition build on the use of indeterminacy in avant-garde music. Sandred et al. note that computers were able to speed up this process, functioning as a “tool that could take random decisions very quickly” (Sandred et al. 2009). While the first three movements of the Illiac Suite were composed algorithmically, based on traditional rules of tonal harmony, the fourth movement was based on probability tables, or Markov chains. They utilized the ILLIAC computer to produce intervals for each instrument, constrained by the “rules” or expectations of tonal harmony (Ames 1989). Following the first performance of the suite in 1956, news sources derided the piece, claiming the audience was “resentful” of being made to engage with an “electronic brain,” with one listener warning that it “presaged a future devoid of human creativity” (Funk 2018). Hiller, in response, emphasized the conceptual nature of the composition, but not without comment on the human capacity for creativity and the socialized constraints of human tastes and perception; in his words: “These correspondences suggest that if the structure of a composition exceeds a certain degree of complexity, it may overstep the perceptual capacities of the human ear and mind” (Funk 2018).

The discomfort felt by these audience members in experiencing something created by an “electronic brain” or “intellectual machine” was most likely triggered by a certain amount of fear surrounding the role of humans in creative endeavors. The reactions align with some of the fears and myths surrounding other AI discourses at the time, during the rise of algorithmic and AI discourse in the mainstream conscious. As Simone Natale and Andrea Ballatore have found, in their discursive analysis of AI in scientific trade magazines, many of these myths and fears solidified during the rise of AI in the 1950s–1970s and continue to inform public opinion and perception of the role and potential of AI technologies. These fears, centered on depicting new computing processes as “thinking machines,” foster skepticism and criticism of AI capacities, as spread and perpetuated through narrative tropes. Generally speaking, the AI technology myths tend to humanize the technology, often connecting AI to ideas of superhuman or supernatural powers (Natale and Ballatore 2017). Terms such as “thinking” and “intelligence” imply a human universality in the definition of consciousness, and AI technologies are thus often perceived to occupy similar forms of intelligence, and by proxy rational thought. The knowledge that this music was not created entirely by human minds seems to evoke a degree of unease, not unlike the response seen in the uncanny valley of robotics and computer-generated human images. Masahiro Mori coined the term uncanny valley to define the unease often reported in response to almost-human images, whereby the small discrepancies between reality and expectation are highlighted to the point of provoking unease. In the case of Natale and Ballatore’s research, it is the knowledge that a computer can create music that causes the unease. In both cases, the uncanny results from assumptions about what is and is not considered human, and the presumed exceptionalism of human capabilities. Even though the Illiac Suite is not considered AIPM, the allusions to the uncanny, the unfamiliar, and the, at times, resentment of computer-assisted composition holds true across genre distinctions.

Some have noted that the Illiac Suite lacks the “journey” necessary for emotional engagement—that there is no melodic drive towards a climax, or overarching sense of purpose. It is difficult to determine whether those critiques are filtered through the knowledge of computer involvement, or if the same could be said of other human-created atonal or aleatoric musical works. Because of the way in which Markov chains function, in that they make predictions from instance to instance, or transition to transition, without the need for extensive information of preceding data, the music created in the Illiac Suite could be seen as a reflection of that form of creation—and thus lacking a “story,” something more akin to a string of independent variables. The music in the fourth movement moves from thought to thought, but does its abstractness belie a sonic narrative. Furthermore, how necessary is that narrative to audience engagement? Avant-garde music continues to push the boundaries of conventional aesthetics, and, in many ways, the addition of the use of a computer is no different. It is interesting to see, therefore, how these technologies are taken up in popular styles, such as Hello World, whereby repetition is valued more so than extended “narrative” content. Novelty, and the uncanny, is often rewarded in pop music, and repetition is key to solidifying audience engagement, both aesthetically and through market saturation.

Current capabilities of AI and generative modelling techniques in popular music are limited to collaborative tools to aid in the discovery/production of novel sounds, melodies, and harmonization. The technology has not yet reached the point of holistic music composition whereby a fully cohesive and engaging popular music song can be created without human involvement.2 However, the problem-solving AIM technique has been developed into something more analogous to true human-computer collaboration. AI software is not creating, in a holistic sense, but neither is it replicating. The act of synthesizing is probably the most apt term to apply to current forms of AIPM, as a form of incremental creativity. As a technological aid, most existing AIPM software can suggest melodies, orchestrations, instrumentations, based on constraints provided; it is a tool for music production, a collaborator, as opposed to an independent creator. As Benoît Carré has noted—it’s like having another person in the studio to bounce ideas off (Marshall 2018). Some of those ideas just happen to push our ears into unfamiliar sonic realms.

2. Flow Machines: Birth of SKYGGE

In 2016, the Sony CSL Research Laboratory, where Flow Machines is located, was credited with creating the first complete AI popular music track, “Daddy’s Car,” inspired by the early music of The Beatles. The track, while interesting, did not spark much mainstream interest, beyond those interested in the process. The software used, Flow Machines, is also responsible for the SKYGGE Hello World album, a much more musically engaging work, and one meant to blur the boundaries of conceptual and commercialized pop music.

From the Flow Machines website:

The main part of the Flow Machines project is Flow Machines Professional. Flow Machines Professional is an AI assisted music composing system. Using this system, creators can compose melody in many different styles which he wants to achieve, based on its own music rules created by various music analysis. Creators can generate melody, chord, and base by operating Flow Machines. Furthermore, their own ideas inspired by Flow Machines. From here, the process will be the same as the regular music production. Arrange them with DAW, put lyrics, recording, mixing, mastering etc.… Flow Machines cannot create a song automatically by themselves. It is a tool for a creator to get inspiration and ideas to have their creativity greatly augmented. (Flow Machine—AI Music-Making n.d.)

The use of the term “augmented creativity” is intriguing, as it centers the idea of the software as collaborative tool, as opposed to entirely AI-created works.

Flow Machines is an evolving project, based on the initial research made possible through an ESRC (Economic and Social Research Council) grant to study potential uses of AI in music production. François Pachet, Pierre Roy, and their team have made incredible strides in the use of various AI systems, most notably Markov models, to create software for AI-human pop music collaborations. Before turning to an analysis of selected tracks from the Hello World album, I first briefly outline the evolution of Flow Machines, from The Continuator, a device to encourage “flow” in music creation, to Flow Composer, a constraints-based lead sheet generator, to Flow Machines and the first AI-human AIPM collaboration. While the evolution of these applications and techniques can be found in a series of publications by Pachet, Roy, and members of their team, a succinct summary of such development has not yet been available, and is outlined below.

The basis for all three iterations of the technology involves a triangulation between an exploration into the use of machine learning in popular music production, Mihály Csíkszentmihályi’s concept of flow (Csíkszentmihályi 1996; Csíkszentmihályi 2008), and the manipulation and development of individual style as an innovative marker of creativity. As such, The Continuator (circa 2004) was one of the first interactive intelligent music systems intended for mainstream users, including children. As Pachet notes, “Unfortunately, purely interactive systems are never intelligent, and conversely, intelligent music generators are never interactive. The Continuator is an attempt to combine both worlds: interactivity and intelligent generation of music material in a single environment” (Pachet 2004). The device continues to play after the user inputs (plays) a musical phrase. The Continuator outputs in the style of the last phrase input by the user, and thus can be used as a tool for music generation. For example, one could play a riff, and allow the Continuator to continue to play in that style, while one is then able to play with/alongside the new materials. It may help to develop improvisatory skills, or function like an intelligent looping system. Pachet’s empirical tests found that children, especially, connected quite easily with the technology, often exhibiting “Aha” effects much quicker and more easily than adult users, and could more often be observed entering into a flow state, defined by Csíkszentmihályi as a mental state for optimal productivity. The device demonstrates intriguing uses for live performance, but also great potential for music education through the potential for scaffolding complexity of tasks.

Flow Composer (circa 2014) functions as a “constraints-based lead sheet generation tool” (Papadopoulos et al. 2016), whereby the application serves three purposes: Autonomous generation of lead sheets, harmonization by inferring chords based on a given melody, and interactive composition. Combining Markov modelling with constraints based on meter solves issues in previous Markov generative models in that it allowed for more control over the music generation process, and more aesthetically pleasing results. By creating models based on a series of constraints (meter, melody, etc.), Flow Composer and Flow Harmonizer are able to generate more successful outputs. Within the autonomous generation capacities, lead sheets are created based on the style of an indicated corpus by first training the Markov + Meter models for chord and melody generation, and applying parameters such as number of chord changes/notes (François et al. 2013).

Flow Machines (circa 2016–present) are a series of “new generation authoring tools aiming to fosters users’ creativity.” Like Flow Composer, they are interactive programs that encourage manipulation and “play” with musical styles. Flow Machines builds on the Harmonizer and Composer, improving upon the generative models used, including text and audio generation. Users are encouraged to “manipulate styles as computational objects” (François et al. 2013), thereby making transparent both the building blocks of creativity, as well as the compositional process itself. The statistical properties of music are arranged into probability tables, as the basis for new generation within certain style parameters. As noted above, this is, for the most part, no different than human music generation, but in human composition these processes occur more often in a subconscious level of perception.

Most popular music is created in adherence, or at least in allusion to, certain constraints or structures, and creativity emerges based on the development of an individual or authorial style within those constraints. Flow Machines draws attention to the function of style in music generation. As Pachet, Ghedini, and Roy note, they “consider style as malleable texture that can be applied to a structure, defined by arbitrary constraints,” and “applying style to well-chosen structure may lead to creative objects” (Ghedini et al. 2015). By inserting style into Csíkszentmihályi’s influential model of creativity, they argue that creativity doesn’t occur in individual works of art, but across a series of applications of an individual’s style. Style—as an extension of skills acquisition—and flow therefore function in tandem to demonstrate the historiometry of creativity, as opposed to its singularity. Just as this proposed model of creativity recognizes the development of style as central to creativity, Flow Machines technologies explore the application of style to musical structures and constraints, as a production of creativity. This begs the question of whether style is something that emerges from lived experience, and therefore indicative of authorial expression, and/or skills development, and therefore a product of probability.

The addition of intelligent musical systems adds a posthuman dimension to creativity, which will be explored in future work. What should be reiterated, however, is that this posthuman creativity is not solely the product of computational models; there remains a human element, both in the corpus used, and the collaboration between human and computer. It’s posthumanistic in its extension of what creativity entails, and the anthropocentric notions of creativity that endure. It challenges the notion that music and creativity are both exceptional and “human,” and deconstructs traditional understandings of these concepts. As previously noted, prior instances of AI and AIPM performances have elicited mixed response, but always with a theme of apprehension and fears about the loss of humanity with the introduction of computers, under the assumption that extending notions of humanity into the posthuman is inherently negative. Using AI as a collaborative tool may help redefine models of creativity, and the expectations of familiarity in popular music production.

Pachet and Roy’s work addresses a lack of holistic approaches to intelligent music generation. Building on “Daddy’s Car,” Hello World represents the first viable application of the Flow Machines applications, and potentially commercially viable AIPM human-computer collaborative album. Beyond its notoriety as the first full AIPM album, Hello World is distinct in its combination of novelty and unfamiliar sounds, within Top-40 pop music conventions/constraints. The following section examines two tracks from the album: “In the House of Poetry” and “Magic Man” as examples of the uncanny audio valley and authenticity of voice.

3. Audio Uncanny Valley: Novelty and Posthuman Sincerity

Because of the way in which AIPM is generated, there is great potential for unfamiliar, novel, and at times uncanny, sounds to be created, not unlike the early uses in AIM works as the Illiac Suite. Human ears, and aesthetic expectations do not influence AIPM generations, but rather they are formed from probabilities, code, and constraints. While it is the prerogative of the human collaborator to pick and choose which musical elements are ultimately utilized, and whether those elements should be manually manipulated, SKYGGE often dwells in that sonic unfamiliarity, using it to their aesthetic advantages. I propose using the audio uncanny valley as framework for discussing these moments.

Introduced by Masahiro Mori (1970), the uncanny valley, as a theory concerning robots and virtual character design whereby closeness to human likeness evokes a negative human reaction (Schneider et al. 2007), is widely contested and lacks empirical validation. Arguments have been made suggesting that there is a generational element in regards to how people react to these uncanny humanlike depictions, marking a difference between those who grew up immersed in virtual characters in comparison to those who did not (Newitz 2013). The same can, or will likely be, said about any notion of an audio uncanny valley. The validity of concept is of less importance than a method for describing the reaction that often occurs in response to AIM and AIPM whereby listeners express discomfort at the addition of a computer “voice,” or the lack of “soul,” or other anthropocentric elements of human creativity.

These reactions mirror those found by the AI myths and fears studies, other early AIM performances, and other outputs produced by Pachet and collaborating musicians. For example, in response to a 2018 concert of experiments in machine learning for music creation, amongst the positive reviews were those who expressed concerns about the new technologies and their effects. For example, one audience member remarked “The computer generated pieces ‘miss’ something—would we call this ‘spirit,’ emotion, or passion?”, with another adding “I think the science is fascinating and it’s important to explore and push boundaries, but I’m concerned about the cultural impact and the loss of human beauty and understanding of music” (Bob et al. 2018). There is a lack of available reactions to the Hello World album to make similar comparisons, and further research should empirically explore responses to AIPM, without listeners having prior knowledge of the mode of production, as that knowledge filters one’s expectations, often looking to “best” the computer in some way or another. Similar reactions have been found to other posthuman forms of musical performance, such as holographic performances of the deceased alongside corporeal musicians, including the 2012 Tupac Shakur Coachella Music Festival event.3 Ken McLeod notes that these holographic performers often “evoke a sense of Freud’s notion of the uncanny in that they manifest a virtual co-presence with both a visual and an aural trace of a larger creative power” (McLeod 2016). He finds that “audiences are paradoxically attracted to the familiarity of the hologram while simultaneously being repulsed by its seemingly artificial, trans- or post-human unfamiliarity” (McLeod 2016). In both cases, a mind/body or human/machine duality is upheld as “normal,” familiar, and desirable, not unlike Freud’s original notion of the uncanny as an interrogation of the alive/dead binary, through one’s reaction to seeing something alien in a familiar setting, such as a prosthetic limb (Brenton et al. 2005). The uncanny, when embodied in monstrous form, represents an “abomination who exists in liminal realm between the living and the dead, simultaneously provoking sympathy and disgust” (Brenton et al. 2005). I argue that the uncanny, in sonic form, blurs the boundary between human and machine production, at times provoking wider fears about the future of human-technology relationships.

Whether or not music can be said to have a “soul” or “spirit” that manifests through human creativity is ultimately beside the point, as audio uncanniness (as a model) can exist whether produced humanly or computationally. The concept of an audio uncanny valley was first suggested by Francis Ramsey to articulate responses to simulated sound, and the perceived “naturalness” of said sounds. The closer simulated sound comes to natural sound, the more human ears pick up on the subtle differences that mark it as strange, or uncanny. As Winifred Phillips echoes, “It sounds almost real…but something about it is strange. It’s just wrong, it doesn’t add up” (Phillips 2015). Comparatively, Mark Grimshaw (2009) takes up the notion of an audio uncanny valley as a positive aim for certain formats, particularly when provoking fear in horror games. He suggests that the defamiliarization that occurs through distortion of sound whereby it still retains elements of naturalness can be exploited to evoke desired emotions. This can be observed not only in video games, but also horror films, contextualizing visual elements. Brenton, Gillies, Ballin, and Chatting note, however, note that context and presence are important; the uncanny audio succeeds through the perception of co-presence. The uncanny valley may be “analogous to a strong sense of presence clashing with cues indicating falsehood” (Brenton et al. 2005) in virtual environments; similarly, the audio uncanny valley in video games and horror films formats exploits those falsehoods to deliberately incite fear and dread. AIPM, however, extends this theory into the purely auditory: Audio cues that lack “presence” in an embodied sense, where the unease emerges from unfamiliarity and unexpected compositional techniques. In some ways this harkens back to the unease experienced in the first instances of schizophonia with the introduction of recorded sound.

In a video describing her SKYGGE collaboration, Kyrie Kristmanson notes: “For me, it is like a folk song, but virtual, digital. It doesn’t feel like a human creation. It is a machine folk song tradition. I find that most interesting.” Furthermore, that “You can somehow feel the melody was not composed by a human” (SKYGGE Music 2017). The idea that a machine could have a folk song tradition is fascinating, as it places the machine within the lineage of the corpus from which it generated new compositional materials, yet she finds it to be distinct from such traditions. The eeriness of the track is undeniable, but it also owes much to the otherworldly quality of Kristmasnson’s singing voice itself. In this track, the piano, strings, and some voice stems are generated by SKYGGE, all instruments are played by SKYGGE (except some of Kirstmanson’s vocals), and is generated based on a corpus of folk ballads and jazz tunes (Credits: Track by Track n.d.).

“In the House of Poetry” is split into two sections—verse and chorus—with the chorus evoking the “enchanting charm of ancient folk melodies” and the verse aligning more with jazz conventions. The Hello World website notes that Flow Machines (SKYGGE) proposed an “unconventional and audacious harmonic modulation” in the chorus within “an ascending melody illuminating the song.” The harmonic modulation is definitely unique within Top-40 standards, but notably, so is the melody (see Figure 1). It is uncommon for mainstream pop tracks to have such a wide range in the melodic contour, adding to its aesthetic of unfamiliarity. What is not immediately obvious to the ear, is that in the second section of the track, the vocal line is generated by SKYGGE from recordings of Kristmanson’s voice (Credits: Track by Track n.d.). For myself, these moments are the ones that are most likely to cause uncanny unease (or even excitement), due to the knowledge of their production. That vocals can be generated by SKYGGE that are largely indistinguishable from Kristmanson’s recorded vocals is remarkable, and hints at future possibilities of generating vocals that are outside the bounds of human possibility, yet sound almost “natural” enough to reach the audio uncanny valley state. Currently, this can be accomplished to some degree by synthesizing the human voice, but the added dimension of autonomous generation may further push these aesthetics into the unfamiliar and uncanny. Through Auto-Tune, and other voice manipulation software, listeners have become familiarized to posthuman voices.

Similarly, “Magic Man” is generated based on a corpus of 1980s French pop music and many of the uncanny elements come about through the nonsensical lyrics generated by SKYGGE whereby the generated vocal line comes close to, but is not quite English. Benoît Carré describes the production process:

The melody and the chords were generated from a French pop song. The melody came very fast. Then, I chose to make Flow-Machines generate a folk guitar and many voices. Flow-Machines adapted the folk guitar track to the chords of my new song and the voices were adapted to the melody. It can be like grafting the coat of a bear on the back of an elephant—but sometimes it gives very nice results! All the generated voices gave a very nice disco mood, using the recurrent word, ‘Magic Man.’ I kept these vocals as the machine generated them, and then added real vocals to get a more precise sound. (Williams 2018)

Of the lyrics, Carré notes, “The machine generated all the unusual and strange lyrics. Listening, it felt a bit like going back to childhood, when you don’t understand all the lyrics of the songs you listen to” (Williams 2018). The fact that SKYGGE generates a vocal line approximating human language is quite uncanny and/or exciting. The repetition of the phrase “Magic Man” draws attention to the lyrics at each repetition, while the remainder of the syllables fade into more subconscious levels of listening. It is easy to casually listen to this track and not immediately realize that the majority of the lyrics are not, in fact, “real.” Just as mainstream music listeners have become more accustomed to posthuman voices, these generated lyrics follow a similar trajectory in regards to the increased familiarity that Western audiences have with other-than-English lyrics. For example, the current rise in popularity of K-Pop in the West has broken through previously held barriers to entry based on language. It is therefore not difficult to conceive of a commercially successful track based on random syllabic generation that approximates the English language.

We know someone through their voice—the connection between voice and embodiment is well established—yet “Magic Man” disrupts that connection. It’s a voice without an origin, a language of only signifiers. I question how that might impact the reception of voice and the perceived authenticity or sincerity of autonomously generated voices. Is it a new form of posthuman sincerity? Scruton suggests that a person demonstrates their understanding of language through its use, in the same way that a musician shows their understanding of the semantics of music by playing (Barker et al. 2013). Through casual observation in playing the track for my undergraduate students, and fellow colleagues, the lyrics for “Magic Man” come close enough to English that, on first listen, the differences are not immediately clear. The uncanniness emerges in repeat listens, as the differences between English and not-quite-English become highlighted, and one’s mind wanders to the further potential of autonomous language generation.

As noted, familiarity with disembodied voices is not new, as they occur not only in recorded sound, but also loudspeakers, phones, radios, and so on. In addition, a disembodied voice alludes to the supernatural or mystical, connecting back to those myths and fears of “thinking” machines. This can be taken together with the fact that in pop music, a completely unaltered “natural” voice is quite rare. In “Magic Man” we have a disembodied voice, an uncanny simulation of voice and language, and perhaps a form of sincerity that emerged in the confusion between disembodied voices as supernatural/authoritative.

4. Conclusions

To conclude with a bit of futurology, I anticipate that in the coming months and years, the use of AI in music production will become mainstream and routine, utilized not only in popular music production, but also on social media platforms, extending notions of everyday creativity and digital sociability to AI. The ways in which these technologies can speed up production, mixing, and even composing, make them invaluable to the production of music, and especially pop music with its culture of ephemera. They can help push human creativity into novel and at times uncanny audio spheres, and this begs the question of how, or if, tastes will adapt. As a tool for music production, AI will no doubt be commonplace soon, but as a creator of pop music melodies, structures, and instrumentations, it is less clear how audiences will approach that content if it sounds too far removed from expectation, or established structures.

In both examples discussed here, the voice is disembodied through AI (re)production and/or “creativity,” which has the potential to affect a certain degree of unease in listeners, often more-so once they are aware of the computational/autonomous production techniques. The elements of aesthetic and linguistic novelty that are produced within AIPM align with a longer history of valuation that is tied to novelty in the Top-40 charts, yet the knowledge of non-human production connects the works to prior concepts of unease experienced with almost-human robotics, as observed in the uncanny valley. The audio uncanny valley, I argue, walks the line between unease and excitement by increasing the potential for novelty, while simultaneously challenging assumptions concerning anthropocentric notions of creativity. Because these practices are in their infancy, there is bound to be a moment of upheaval before they become mundane, and mainstream familiarity with computational creativity is increased and “normalized.”

I am most excited for what becomes of the unexpected uses of AI content production. I anticipate their incorporation into mobile phone apps, whereby one can easily create custom music to accompany videos, photos, etc. Or infinite “chill” playlists in that are entirely AI produced and platform-owned. The future of creativity and humans’ role in that future is speculative at best. Regardless of the “intelligence” of AI, it is principally human-driven and consumed, and, as such, it will be human agents who ultimately guide its use and progress. SKYGGE’s Hello World is a product of these new forms of production and consumption, and functions as a pivot moment in the understanding and value of human-computer collaborations. The album is aptly named, as it alludes to the first words any new programmer uses when learning to code, as well as serving as an introduction to new AI-human collaborative practices. Hello, World, welcome to the new era of popular music.