GHOST IN THE MACHINE: SUNO

In the first article in this series, my co-author, Dr. Charles Eldering, broke down the immediate copyright issues the AI music industry currently faces. As it turns out, that may be the tip of the iceberg. Read the last article on Light Drafts

While Charles holds down the research on the technology and legal fronts in AI music, I’ll be focusing on the perspective of its sound itself, whether it indeed does what is advertised, from the perspective of a veteran of music production, and to determine what do these tools generate when used in either a relatively unguided mode, focused on text with emotional dynamics, to find ways of how it can be also used as a tool within the creative process, in a collaboration where AI plays a role yet the human remains the main creative conductor and director. For now, in this episode, we will keep to the text and lyrics process within the Suno model.

Forensic and Creative Studies of Popular AI Music Generators

In the cases around generative AI music, gathering evidence sufficient to convince a jury of copyright infringement, false endorsement, or other alleged claims requires expertise in several fields: audio and digital forensics, acoustic forensics, musicology, and, in some cases, ethnomusicology. The digital tools for this work are either common or hard to acquire, as they are mostly reserved for the enforcement of legal claims. This will not stop us from using creativity and some old-age wisdom gathered from individual and community experience in the fields, to figure out what is going on under the hood of these generators.

Share

Forensics: The application of scientific methods, technologies, and investigative techniques to establish facts, recover hidden data, and analyze evidence for legal, diagnostic, or investigative purposes.

Dub: a historically Jamaican-rooted musical genre and avant-garde studio and creative philosophy developed by sound engineers centered on using the mixing console, reverb, phasers, and echo as a primary instrument to dismantle and recreate an existing recording.

Years ago, at the Underground Producers Alliance, I developed and released an EP, which was both a music production release and a tool for practitioners studying sound system performance, engineering, signal processing, sound design, and futuristic production methods, called “Whine The Leaf” (as Ghost Producer).

The vocal version of the song, PSA featuring K.A. and Tori Nicks, was later licensed and used as a score by fashion brand Prada for their fashion shows in Asia and Italy, however, unbeknownst to anyone accept for students, colleagues and inquiring fans, a bonus track yielded a lesson inside of a .Wav file’s spectrogram around phase, noise, equalization, resonance, mixing, mastering and spatial sound. Each did a different version (aka Forensic Dub) of the analysis file. Every student who participated was later far more advanced, not only in correcting those issues but also in hearing them in the first place. Zip forward today, a lack of critical listening works spectacularly for those currently being sued, but cases have been won on style alone, so it is not set in stone either.

A less technical and more obvious example, to me at least, of this apparent use of non-critical hearing would be if you take into account that perhaps a musician has been uploaded into AI models who may not be of Western origin at all, but is clearly coming out the other side of a generator. It would currently be very difficult to prove, in an American court at least, because, apparently, it is ok for the layman to know the difference between Rap and R&B music but not hear the difference between music from Morocco or Tibet, even when a layman from either of those countries probably would almost certainly know the difference. It is due to that much-relied-upon ignorance that musicians may not be able to win cases, especially if their evidence is within lesser-noticed but real and important aspects of sound and music.

Putting aside the sound of the music itself, the Forensic Dubs purpose is to accelerate sonic awareness around the processing, effects, and other left-out aspects of a musician’s output. It is about pushing the limitations of sound systems and instruments. Jimi Hendrix was not just with a guitar; his experience included a Marshall Amp, a 1/4” cable, processors, a microphone with yet another cable, and an acoustic space captured on obsolete media with artifacts. That is the reality of the entire experience, and yet it is vastly ignored. AI Models are just as ignorant as their human teachers, and one can throw as much incomplete data around as they like; they will never arrive at the right conclusion. It is also here where a similar approach of Forensically Dubbing would be useful, but to do it with AI Models like Suno, the Dub is inverted, it is no longer an instrumental, it is not a vocal either, lyrics and text inputted into an AI Model with little capacity to experiment otherwise are what makes it sing and tell us everything that we need to know, bringing us back to the forensic part of the Dub.

Whine the Leaf Analysis File (FREE DOWNLOAD)

Some tips for users, in my testing, I found it was important to utilize the lyrics panel to design the changes in the music it generates. For example, users should experiment with symbols, code, and even just brackets to see what happens. | & \ can go a long way, but sometimes will do absolutely nothing.

The text used for lyrics is either written by me, some of which has been published in various places, or borrowed material from accelerationist philosophers (AI and Nick Land are like peanut butter and jelly, as he planned). In one instance, I threw in an article I was working on about Phonk music and added “Phonk” as the genre. It seemed to like that.

You can also hear my Memphis Phonk Oral History Suno Lecture here. I tweaked it a bit… This is a collaboration between the lyrics and the styles; both cooperate, and due to the text being about Phonk, and the style being Phonk, the LLM knows what to do with it, that is what these generators are built for, clear appropriation of formulaic aspects of music, well packed into data sets.

Memphis Phonk Suno Lecture | Prompt: written by Raz Mesinai

Here is the entire playlist from my tests on the Suno platform.

Although I developed this playlist on prompts alone, a lot can be done with uploading your own material, which involves true human contribution, which is badly needed here. However, the legal challenges the AI companies face are causing constraints that can hinder the artist’s ability to do so, and thus create an obstruction in the creative workflow. If we consider how sampling in music was done before platforms like Splice emerged, there was a direct connection between the sound being sampled from, say, a vinyl record, the plugs and pre-amps or mixer to an MPC or other hardware sampler, and at the very center was the producer, with whatever eccentricities came with that individual.

In addition to revealing the degree to which AI music generators have been trained on well-known and lesser-known artists, Forensic Dubs serve as a litmus test for the quality and authenticity of the model’s output. With a little patience and careful listening, even a previously untrained ear can be trained to hear the degree of “sameness” as well as unique characteristics between tracks in each of these platforms.

After prolonged play of the playlist, one can hear that all of them currently hold a very clear sonic identity. This can become a big problem for some who pronounce their Model as a DAW (Digital Audio Workstation), but fine for those using it as an instrument, which is more my line of thinking. On the other hand, there was a time when Ableton Live was said to have “a sound” as well, and it is still one of the most popular. MAX MSP also has a very apparent sound, and it is indispensable for sound design.

Here is another example, where I have taken one of my articles on improvisation and used it as lyrics once again. The flow is quite well put together, but the music is nothing close to improvised, which is to be expected since this is Suno, not another model like the ones we will explore in upcoming articles that can produce such material.

IMPROVISATION AND SELF-DETERMINATION SUNO LECTURE

Share

I believe that over time, it will become abundantly clear that most of the popular generators, like Suno and Udio, are built only for mainstream pop and dance music culture in mind, the true essence of sound system music is quite lost as refined data due to its complexity, but it could be if experts were brought in, and, clearly from the sound, they weren’t.

If one attempts to get experimental, the platform will almost always correct things, as it only knows what it knows, which is limited to mainstream music and its close derivatives. For experimentation, however, some new platforms will undoubtedly emerge; the temptation and demand are too high, and these will be the most useful for generating sound designs, accurate clones of regional instruments, voices, and more complex musical generations.

For them, experimentalism in a genre comes after a genre has been formed. That is incorrect. The genres are first developed from an experimental approach to sound, instruments, or a previous genre. The experiments came first, and it is experimentation that allows a genre to innovate and survive later.

Again, I think AI will offer musicians new tools for their craft that allow music to continue its evolution as a reflection of humankind’s history, struggles, and aspirations, but simple prompt-based music generation is not one of those tools, unless experimentalists build a theory to cause hallucinations on purpose, mastering that, and owning the control of beaking and fixing it till it sings from a human standpoint. Hacking is inherently human and is thus an art.

For next time: There is a darker side to AI music: The physiology of reward and the role of music in dopamine and anticipation is nothing new; it’s well studied and documented.

Could AI music be used to elicit effective responses from listeners and potentially manipulate them?

In the meantime, enjoy (or cringe to) this hot new number by 50(NON)Cents and his band of imaginary friends, done in a different model, TBA.

Share