Steal like an AI — How AI might generate music archetypes
Exploring artificial intelligence as the perfect musical accomplice.
Every music artist under the sun knows that originality is a lie. Whether you call it borrowing a style, drawing ‘inspiration’ or covering a riff, there’s no two ways about it: you’re repurposing someone else’s work as your own (or exchanging ‘musical memes’, as Steven Jan puts it). And that’s not problematic; if anything, it’s the scope of what you can get away with that says if you’re skilful. (The less obvious your influence, the more credit you certainly get). But what happens when you employ an AI to assist in your ‘re-creation’? Perhaps this is the workaround that producers who can’t afford expensive sample clearance or feature fees so desperately need.
Symbols, waveforms and features
Bringing your dream feature to life — a collaboration with an artist, producer or musician you view as iconic — is a frequent motivation for artists-in-the-making. Many start out admiring the musicians they look up to, and some even get to work with them later in their career. Thanks to advances in AI, you might be able to give your song the touch of John Lennon you’re looking for. Here’s how it works.
Broadly speaking, you train an AI on the data you want it to learn from. Feed it the entire Beatles discography, and it could potentially learn how they write and play, and hopefully, how to produce songs in their style — at least, that’s the long term vision. Creating the sense that Lennon and McCartney might’ve sat in on your songwriting session could be achieved in varying ways. Some approaches are more useful for certain cases, such as composition, or lyric generation (although this is still a stretch). Perhaps the most musically conducive approach is symbolic analysis.
There are several reasons why AI music experts are fans of this approach. First, it delivers the most musically rich information. Each score contains important characteristics and structures, including harmony (chord progressions), theme, and instrumentation (including the role of each instrument and how it’s played). You also find out how the music is ‘shaped’, meaning the injection of intervals, dissonances and ornamentation. Ornamentation includes trills (fancy, skilfully played alternating notes), arpeggios and other techniques gloriously executed by the likes of Chopin and others.
The easiest approach is to start with the chord progressions of a piece of music. Unsurprisingly, this is simpler if you’re in the realm of pop music and only have four chords to consider. You’d then examine the sheet music, much of which is freely available online, on sites such as MuseScore. But there are also some types of music that simply aren’t notated, such as traditional folk music. Based on an oral tradition and passed down through the generations, it originated in a time period where the tools for notation were non-existent and songs required memorisation. This proves challenging for those specifically interested in the applications of folk music in machine learning, such as researcher Bob Sturm who recently published an article detailing his work. Gathering the symbolic information from unrecorded, often antiquated music proves painstaking; it requires manual listening and capturing of the musical elements such as chords, pitch and rhythm information into a computer. The faster way might be to simply program the computer to capture this information automatically.
At the other end of the spectrum you’ll find waveform or audio analysis. Let’s say you have a specific song in mind that you’re looking to replicate. You could implement signal processing techniques — of which there are many, such as RMS (root mean square) calculation or spectral analysis — to automatically identify the progression of chord changes in the audio file. As you might’ve guessed, the downside is that a waveform doesn’t carry all the musical information you’d find in a score. Compared to extracting information from a symbolic score, it’s difficult for an algorithm to detect the rhythm, harmony and musical variations in an audio file like an MP3. On the other hand, it contains fully-formed sound; the final rendition of all the instruments and the way they’re played. This means the algorithm can learn to detect the unique performance style of the artist, the genre, and timbre. The ability to scan each of these components means you get a closer imitation of your would-be collaborator.
Both of these techniques rely on feature extraction; basically pulling numbers out of any data, or even a musical dataset. This works for any features you select, no matter the discipline: density in photos, event detection in video games, or ‘low-level’ and ‘high-level’ features in music. Low-level features, for example, refer to the loudness — the overall energy of a track — or how much bass or high end is present (spectral characteristics). ‘High level’ features are broader in referring to the musicality; think ‘danceability’, or the type of genre it fits into. As algorithms process the combination of these core features, they inch closer to replicating what’s essential about an artist’s style.
A case for stolen identity
The Russian-born composer and piano conductor Igor Stravinsky famously stated: “Lesser artists borrow, great artists steal.” (Ironically, this phrase is also attributed to Pablo Picasso; perhaps it was stolen, too). If becoming a great artist means stealing influences from others, then why not use the latest tools to do it. Since the dawn of sampling, music technology has helped producers pay homage to their inspirations directly. But the price and hassle of sample clearance — the process of paying the owners of a piece of work to reuse a song — can become draining, or even impossibly expensive. (Maryland rapper Logic’s recent outcry emphasised the many hurdles producers face.) You could always opt to link up with the artist you’re inspired by. But it comes at a cost. Once an artist reaches a certain level of fame, so does the fee for a feature. (Rapper 2 Chainz claims to charge a hundred thousand dollars to rap on a song). And either you pay up, or at least offer unique creative input that makes you a worthwhile collaborator (which should always be the minimum, unless you’re one of those musicians playing for beer and pizza).
Artists wear their influences with tremendous pride, often generating some of their best work off the back of inspiration from what they’ve been listening to that week. The Rolling Stones’ smash hit ‘It’s Only Rock N’ Roll’ was described by Mick Jagger as very “Chuck Berry”. Kanye West, and other hip-hop pioneers like Q-Tip from A Tribe Called Quest have built entire catalogues by sampling directly from others. But the point is not to blatantly borrow without credit, or skill. As Jagger adds: “You always have to start out by imitating somebody…and then you develop your own (style)”. The point is to steal as much as you can, and then give it your own spin. If you’re not as skilled, you might find it harder to replicate the style of the musician you love with a certain level of authenticity.
That’s where AI can come in. It’s cheaper, and can systematically learn what makes the artist you’re drawing from uniquely ‘them’. It democratises access to your dream collaboration across time and history, provided you’re willing to find the data and learn how to program it (or find someone else to). The result could become a more authentic impression than anything you could ever create. But the danger of distillation is that a well-crafted piece becomes a shadow of its predecessor; the barely-recognisable, mutated bones of the anthemic music you admire. Although, as the technology advances, the accuracy of its representation will surely magnify. Still, this may provide an effective, strategic tool to capture your favourite artistic inspiration into a series of archetypes you can pull out when you need, like a team of imaginary famous musician friends. Talking to them is entirely optional.
