VULANE MTHEMBU & HEIKKI SOINI
Nguni Machina remixed
Part 1: Nguni Machina: Explorations in AI, neural networks/machine learning music generation
by Vulane Mthembu
The initial motivation for the Nguni Machina project was driven by curiosity: How compelling can AI-generated music sound in 2021? As this was in itself a very subjective question, without a clear-cut answer, I had to establish some form of yardstick to gauge the success of the project.
This is where I turned to the age-old Turing Test proposed by one of the earliest figures in Artificial Intelligence, Alan Turing, who himself in the early developments of AI theory had dabbled in rudimentary art created using an early computer in 1951. The Test: a test of a machine’s ability to exhibit intelligent behaviour equivalent to, or indistinguishable from, that of a human. If the art created by the machine is indistinguishable in a blind test from its human counterpart, this would pose some measure of success.
The results were surprising, they were flourishes of “artistic brilliance” in some of the pieces. This led to the development of Nguni Machina into a more expansive project that would not only address AI but other fields of particular interest.
Nguni Machina is open source and released under the creative commons license to encourage interaction and interrogation of its questions.
Who owns art created by AI?
What is the role of a human artist in the era of compelling AI-created art?
How easy is it to create such works and the practicality of the exercise?
The project was created using Google’s Magenta which is an open-source research project exploring the role of machine learning as a tool in the creative process and MuseNet AI, a deep neural network that can generate musical compositions from collected training data from many different sources. using large collections of MIDI files spanning several genres, including jazz, pop, African, Indian, and Arabic styles. Additionally, a MAESTRO dataset was used..
It was important that I did not interfere with the AI in any way nor introduce any human assistance in the complete album creation process from composing and arranging to mixing and mastering which was all done by the machine. This was done to ensure an as accurate as possible representation of what AI is able to create without any human assistance besides the initial data set training.
Also, it was important that the album did not overstay its welcome and remained with a relatively short runtime which ended up being 5 minutes, enough time for the audience to grasp the concept and allow interpretation of its implications. AI-created music is not anything new, as great examples recently by Babusi Nyoni with his Gqom Robot and others have proven to be exciting explorations of what is possible if we let the robots run free in popular culture.
There are two major camps of thought in terms of how AI in its current or immediate future could be:
AI as a collaborator
AI as a creator
Could our collaboration with AI lead to new kinds of art we’ve never before imagined?
In International Conference on Learning Representations, 2019.
MuseNet by Open AI
Payne, Christine. “MuseNet.” OpenAI, 25 Apr. 2019.
This excerpt was initially published on Medium on 22 May 2021
PART II: A Human and a Machine Collaborating
by Heikki Soini
Last year I was asked to make a remix of a song that was originally made by artificial intelligence. A person who asked me this question was Vulane Mthembu and he had just released the first fully AI-generated music album in Africa called Nguni Machina. Vulane knew me and I’m sure that he was certain that I would answer his call. The question from the beginning was more or less about the music. How does music sound when it is made entirely by AI from composing and arranging to mixing and mastering?
I had made remixes for other artists before this, but there had always been a mutual interest in a certain type of music genre and respect towards the artist whom I was remixing. Now I was dealing with a faceless computer. Could I possibly relate to the music enough to make a remix? I didn’t have any relationship with the real artists of this project, which was Magenta by Google MuseNet AI. When I was a teenager sucking musical influences from death metal to progressive rock and jazz like a sponge, tools like Magenta and MuseNet AI were just science fiction.
This is a little embarrassing to admit, but when I was listening to Nguni Machina my first reaction was a relief. I didn’t quite know what to expect and how good AI was at making music. As a silly and proud musician, I had the urge to compare my skills to a machine.
Now when I heard this clumsy music that sounded like it was played by a beginner with cheap midi instruments, I felt more at ease. After finding a nice piece of melody I knew that I could do this.
The original melody that I picked was played entirely by a sound that modulated the cello, I guess. There were no chords or rhythm elements. I ended up sampling just the first 7 seconds of the original track and composed my whole version around that main theme.
I started with transcribing the melody and then built up the harmonies. Because of this foolish competition that I had in my head with the AI, I wanted to make the chord progression as complex as possible and the same thing happened with drums. I can honestly admit that it was all just a show-off – I wanted to beat the machine. The last thing I did was to scratch a vocal line concerning AI from a south African artist called Kwela from another project that we did together. After that, I felt that I was ready.
Did I manage to beat the machine in creativity and is that even relevant by any means, I leave that for you to decide? One thing is for certain and it is that artificial intelligence has come to the arts and music to stay. There is no turning back now.
Nguni Machina: Collaborations
/ NGUNI MACHINA: Open-source neural network explorations in artificial intelligence-generated music. Released under the Creative Commons (cc) license for everyone to freely remix, share, rearrange, destroy or improve as they wish. Published 10.05.21
Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset:
Curtis Hawthorne, Andriy Stasyuk, Adam Roberts, Ian Simon, Cheng-Zhi Anna Huang, Sander Dieleman, Erich Elsen, Jesse Engel, and Douglas Eck.
“Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset.”