To mark the release of her new album Proto, Holly Herndon skypes Zacharias Szumer to tell him about social protocols, digital musique concrète and how it takes a village to raise an AI vocal-bot.
Well, we are now certainly in the depths of the uncanny valley – that gulf somewhere between Blade Runner replicants and C-3PO in which artificial intelligence bots are just human enough to be creepy as fuck, but can’t, for example, take a bus trip without raising eyebrows. Anyone with an internet connection and a morbid fascination with new tech watched on as Hanson Robotics’ Sofia made her debut in 2016 – holding press conferences, making talk-show appearances and briefly dating Will Smith.
Chatbots like Sofia are already pretty good value for conversation, but can you teach an AI to sing?
That’s the question electronic producer and vocalist Holly Herndon set out to explore on her new album Proto (out May 10 through 4AD). And to answer it, she and her collaborator Mat Dryhurst ‘gave birth to an AI baby affectionately named Spawn’.
First things first: Proto is definitely not the presentation of a child reaching maturity – it’s no AI bat mitzvah. It’s more of an artificial-intelligence birth video, with all the blood and placenta on full display. The clearest example of this is the track ‘Godmother’ – a gravelly, stuttering beatbox number that was generated by feeding Spawn a heap of Jlin tracks, and then coaxing her to recreate Jlin’s style in Herndon’s voice. Spawn’s not exactly about to win an Aria for Best Female Artist, but as the first words/song of nascent technology, that’s beside the point.
As Dryhurst explains, ‘“Godmother” is obviously a really ugly piece of music, but that was the point. So many idealizations of this music will sanitize it. This was an illustrative piece of music for where the tech is at. It’s a raw moment … like a baby babbling and mimicking its parents. Humans replicate digital processes; digital intelligences replicate humans.’
To train Spawn, an ensemble of vocalists were recruited to spoon-feed her hours of improvised acapellas at weekly ‘learning sessions’. Spawn, for her part, lapped it all up from her cradle: a ‘souped-up-gaming-PC’.
‘We wanted to put together a community of artists for this,’ Herndon says in the Proto press release. ‘There’s no escaping the hours in front of the computer, but we also craved a very physical in-person sound and the experience of music-making. Platform [Herndon’s last album] was platform-based in terms of its genesis as we were primarily working with people over the internet. I grew up singing in choirs, and I missed that call-and-response and mutual inspiration.’
It was also important for Herndon to acknowledge her collaborators on Proto. The video clip for the lead single ‘Eternal’ is a mash-up of the faces of the vocalists who trained/raised Spawn. Unlike so-called Potemkin AI, which often tries to hide the human cognitive labour behind the machine, Herndon wanted to put it front and centre:
“You can hear the community that trained Spawn. With much AI that we’re seeing right now, [you get]this perfected, glossy output where you don’t see any human labour behind it, and we think that’s a fallacy – we want you to be able to hear the people that went into this training and for people to realize that it is actually a human endeavour that takes many humans. We didn’t want to erase that work …”
As they say, it takes a village …
‘We decided to use a child narrative because – while we’re not seeing her as a human child, or trying to anthropomorphise her – we don’t think that we necessarily have to anthropomorphize intelligences to care about them. We do think it’s a kind of intelligence that deserves the kind of care, and almost the kind of weight, that goes into raising a child. The way that it takes a community to raise a child – that you introduce your ethics and the values that you want your child to have at a very young age – that’s something that we’re not necessarily seeing in a lot of research right now. It’s more just trying to figure out how to make things work as good as possible without really thinking about the protocol layer on which everything is built, that can kind of code in some of the human values that we have.’
The album’s name, Proto, is derived from this conceptual double entendre of a ‘protocol’, which in computer world is a set of rules or procedures for transmitting data between electronic devices. Of course, computer protocols often reflect the power dynamics – the protocols, if you will – of the society in which they are coded. As Herndon says, ‘protocols’ emerged as one of the main conceptual ‘vectors’ of the record from …
“… thinking about how – at a very early stage in technological development – one might design the values and ethics of the specific community that’s using that protocol into the protocol itself. Protocol doesn’t have to be a technical word – it can just be the behaviours that communities agree upon for a certain activity. But [we were]looking at that in technological systems and wondering how that can be addressed from the beginning. I often think of the internet as an example. What is it about the way that the infrastructure of the internet developed that allowed for platform capitalism to run so rampant on top of that initial protocol layer? What if there had been certain attribution layers from the beginning, or privacy, or data self-ownership values hard-coded into the protocol layer from the beginning? Would we be in a different situation? Would Facebook or Google or whoever not be able to just hoover up all of our digital selves and monetise that in an advertising paradigm?”
Indeed, it’s a somewhat grim world that little Spawn is being uploaded into – especially as an AI bot coded as female. The digital age has, as we all know, been quick to monetize toxic masculinity through subservient sex-bots etc. (In this sense it only seems natural that the first country to give Sofia citizenship was an arch-conservative theocracy that has historically done a lot to turn the members of its female population into subservient automatons). As Cher Tan wrote in an article for Matters Journal recently: ‘Both AI and women are based on eerily parallel constructs. In the quest towards authenticity, naturalness or believability, both the woman and machine must behave in such a way that the person interacting with them is completely convinced there is nothing amiss. Little wonder, then, that so much AI is gendered as women, even if robots are inherently genderless … Gender operates in these systems to signal domesticity and trust-making … people feel comfortable having these otherwise intrusive technologies in their most private spaces ….’ I put this as a question to Herndon, who has previously explored servile femininity as a construct in her ASMR boudoir lullaby ‘Lonely on the Top’:
“First of all we refer to Spawn as a ‘she’, and this has been questioned by several people because of the history of subservient AI – helper/server things like Alexa and Siri. For me, it was a matter of starting to train Spawn. The first files that I was training her on were my own voice, so I saw a piece of myself reflected back to her … I saw a piece of myself in her, and that’s why I felt she was a ‘she’. I think that just because some people misuse language or have a bad track record with certain things doesn’t mean they get to own the entire topic. So for me, using ‘she’ was a way of taking back that approach in a way that was an ethic that I feel more comfortable with.”
For most of her musical career, Herndon has sought to blur and subvert a common musical dichotomy – the gulf between the ‘aesthetic of authenticity’ (the privileging of raw emotion, naturalness, ‘realness’, soulfulness, organic/analogue texture and a rejection of effects such as autotune) and the ‘aesthetic of futurism’ (the foregrounding of robotic, cold, hyper-rational emotional zones, cyber-sploitation tropes/sounds, autotuned/vocoder vocals, etc.). Of course, this is a pretty tired dichotomy – the authenticity aesthetic is now too reactionary, and the ‘futurist’ aesthetic has just become ‘an established style, much like a particular typographical font,’ rather than any novel vision of the future, as Mark Fisher puts it.
In opposition to this false disjunction, Herndon has developed a protocol for digital music that accepts augmentation but isn’t heavy with the airbrush, and an aesthetic that isn’t centred on the often-over-exaggerated de-humanising effects of technology. ‘Our vision of technology is that it enables relationships and liberates us to be more human together,’ Herndon has previously said. And she believes that the laptop is ‘the most intimate instrument’ we have. This approach is reflected in the digital palette she uses to construct her albums – ‘net-concrete’ Max4live patches that smoosh together hours of output from one’s internet browser, custom vocal patches and unique live voice processing systems, for example. The output of these artistic choices is a sound that doesn’t cohere with any cheesy, stock futurism. ‘Eternal’, for example, is a heady mix of Herndon’s vocal-collage shtick with the infectious lure of a Mariah Carey love song. In training Spawn, Herndon also took the digital road less travelled, and it has made all the difference.
“Rather than work with automated composing, which is where a lot of AI research is going – studying pre-existing scores, then extracting midi data and creating a canon from which you can compose forever in the style of Beethoven or whatever – we decided to stick specifically with audio material. We found [automated composing]a really boring approach to AI. First of all, because it gets us into an aesthetic recursive feedback loop, where we’re always in some ways recreating what came before, rather than building on it. So we decided to stick specifically with audio material, moving more in a lineage from musique concrète, where we’re actually dealing with sound as material. [With] sound as material, you hear how nascent it is. It’s not this perfect magic wand yet. When people extract midi data and create these automated compositions, they usually push those through digital instruments, or sometimes human performers, and it gives this sheen of perfection to the whole process. When we use sound as material, you can hear the crunchy timbre of the network trying to figure out where the sound should be sculpted next. “
Spawn isn’t human-enough to pass the aural Turing Test yet. But, of course, ‘passing’ as human isn’t the point.
“I think so much research in AI now is towards this kind of accuracy, because that’s what works in the marketplace. If you can create some sort of a system that makes a film score that is indistinguishable from Hans Zimmer then there’s a lot of money to be made there, right? So a lot of people are running head-on towards that. For us, of course, it’s always a push-and-pull. We try not to be too goal-oriented. We want to create a wide spectrum of data sets and see what happens; it’s more of an experimental approach than a goal-oriented approach.”
While Herndon hasn’t created a bot to incorporate Spawn into a live stage performance yet, she’s still put her out into the world at a young age. This seems like a risky move. We’ve all heard about those millennial couples who create an Instagram account for their gestating foetus or document every day of their toddler’s life via Facebook. And that’s just everyday people giving their child a non-consensual digital presence among their 300 or so friends. Herndon is a semi-famous artist! Given the celebrity of AI bots like Sofia, I had to ask: is she worried about Spawn becoming famous too young and her life trajectory following that of nearly every other child celebrity?
“Haha. That’s a kind of hilarious question. We’ve purposely chosen not to give Spawn an avatar because we don’t want to fall into those kind of tropes. We also see it is something that’s constantly evolving and changing, and if we were to give Spawn an avatar, then Spawn would be fixed to that image. So most of the times that we’ve expressed Spawn visually, it’s been more of a morphing of the ensemble members’ and collaborators’ faces together – Trying to make visible the human community that’s training Spawn. Rather than trying to create a celebrity around Spawn herself. To go back to the birth idea, we’re trying to think about how we can, in some ways, de-centre human intelligence as this kind of apex of intelligence. So thinking about ideas around the inhuman and other kinds of intelligence and how we can learn from that, and how we can have relationships and gain from that. So looking at Donna Haraway’s writing on kin; she talks a lot about relationships with animals as kind of way to view ourselves and our position in our home planet – to try to move away from this extremely human-centred worldview, while still maintaining the aspects of humanity that we really value.”
Holly Herndon’s third full-length album PROTO is out the 10th of May via 4AD/Remote Control. You can find it here.