Brave New Sound? We Asked Music Pros to Review AI-Generated ‘Songs’

First AI came for images, then for words, and now it’s coming for… music? While stock music already exists, it also has a bad reputation: generic, forgettable, cheesy. Can AI-generated music fare better? That’s the question we put to some of the industry’s music pros – asking them to review a couple of AI-generated tracks we created. And they didn’t hold back…

In 1888, the essayist and art critic Walter Pater wrote that “all art constantly aspires to the condition of music”, expressing the idea that while other forms of art strive to fuse subject matter and form, in music the two continuously duet.

Music has a way of stopping us in our tracks, taking us back in time or far into the future; it can scare - raising the hair on our skin - or warm us deeply with a sound brighter than midday sun. “Music has the ability to affect us in a way that only drugs, sex, alcohol and religion can compete with. If AI is ever going to replace composers and bands it has to be able to comprehensively understand so much about things that are so incomprehensible” says Simon Elms, director and composer at Eclectic Music and Bark Soho. 

“Character? Music is all about character. Genius and dexterity – two things that computers can do - are also important but not necessarily essential. Johnny Marr has never been regarded as the best guitar player in the world but he’s definitely one of the most influential ones - the way he uses the instrument and the original riffs and sounds that he’s produced over the years have resonated with millions of people and have been the backbone of some huge songs. His playing is original and, most importantly, characterful. Music is about people. It’s about talent, muscle memory, sinew, passion and love.” 

“And then there’s the interaction. The band. Lennon and McCartney. The alchemy of groups of musicians playing together. How can a computer ever compete with that? How can they replicate a recording of Mozart’s requiem recorded in Westminster Abbey with a 100 piece orchestra and a 100 strong choir? The sheer physicality of it is mesmerising. It’s overwhelming but in the best possible way. This is what we’re good at. This is what human beings are all about. Computers are brilliant at research; maths. Why bother composing and producing music? It’s pointless. It’s self-defeating. We just don’t need to do it. The proof of which is below. Putting music like these tracks on your precious imagery is like eating your favourite food and washing it down with Toilet Duck.

To create the ‘songs’, we used the software, Soundraw, and gave it a few prompts (a brief) from the available keywords to generate four tracks for different scenarios. Below, Simon is joined by Gabe McDonough, EP and music supervisor at MAS, and Robert Miller, global creative director and composer at Sound Industries in reviewing the AI’s sound stylings…

LBB Prompt: Ads and trailers, sexy, rock

Track: https://soundraw.io/edit_music?m=63da3c5d8f35b4001cf4d604 

Gabe> This isn't good, quality music by any stretch. But that same criticism has been levelled at everyone from the Rolling Stones to Lil' Nas X over the years. 

The fact that an AI can even create something today that sounds as good as this piece is impressive on some level, though. I'll be even more interested to hear what's getting created five years from now. In the near term, I think that the first mass culture moment with AI music is going to come in the form of a sample that is twisted into something else and augmented by human composition. Not that different a concept than pulling something from Splice really. It will likely be a sample of something that sounds "bad," "weird," or "cursed' in some way-- leaning into the AI-ness of the sounds rather than trying to pass as human. This is already happening in the underground with UK electronic artist Patten announcing the April release of his Mirage FM album. The album was created using text descriptions that AI converted to audio which he then resampled and manipulated to make his record.

Robert> I would like to not seem sarcastic when reviewing AI compositions and simply react to what I hear, but I think it may be hard not to take a rather dark viewpoint of this ‘music’. This first piece's prompts are ‘sexy, rock, etc.’ Okay, these are rock genre programmed samples, but it's completely devoid of the guttural human quality of sex in any way. It's built on two chords and a stiff  ‘guitar’ riff, but that's the least of the issues. Think about how alive two-chord Rolling Stones songs are (as just one of a million examples). The so-called music has as much life and sex as a cheap video game track - and what's even more depressing is the lack of varied expression in the way the chords sound. Music that is created and played by a human involves countless subtle variations in chord strums, drum strokes, and keyboard strikes - all the things an actual human rock performance involves.

Simon> They need to really nail down the definition of each of the words they allow you to brief with. This rock track is definitely not sexy. It relies on five separate four bar loops that all fit together. Interest is just about maintained by adding a new loop every four-ish bars. There’s no interesting lead lines and the production is as basic as it comes. It’s horrible. Please make it stop.