Eurovision 2023: RAIge against the machine 

At AxiCom, I zoom in on the “so what” of technology: its impact on our clients, us, and our daily lives. I do this to help our clients define the role they play in the world and communicate this role to their audiences. But Generative AI reveals a dilemma – how can we as individuals, and as an agency, benefit from AI? What are its limitations and should we embrace the possibilities?  

To better answer these questions, and discover where media hype differs from reality, I have been part of the AxiCom Generative AI Eurovision song challenge. 

My role as the composer was chosen because I have played in bands since I was 14, and have spent many days in rehearsal studios, experimenting with sounds and rhythms. The idea was this would be enough knowledge to produce a musical arrangement (to accompany the AI-generated lyrics) with the help of only free Generative AI tools. 

The sound of musAIc

My starting point was to experiment with multiple platforms. Some are advanced and all you need to let them know is genre, tempo, key, length and “mood”. Easy enough for a Eurovision song for which there is a scientifically proven recipe to follow. Just mix Europop, 120-130 beats per minute, and major chords.

I plugged these parameters into a few platforms such as Mubert and Soundful, and they generated songs with melodies and rhythms. Some were, let’s say, “unusual”. Others were pretty catchy, like this one created on Soundful.

I felt I was on to something…but that is where the magic stopped. And it took me a while to figure out why.

There are some structural conventions in pop music. For example, most songs start with a brief intro and a first verse which leads into a chorus after 35-45 seconds. It then quickly goes into verse 2 and then back to the chorus again, followed by a bridge that creates “drama”, before ending with the chorus. Most pop songs follow a similar structure. But my AI songs had no real highs and lows, build-up, or change in keys. There are plenty of tools for making manual edits, but I could not find a tool that uses AI to add structure and drama. 

Slaves to a rhythm

So, here I was with a few half-satisfying AI-generated songs that lacked structure and emotion. But most of all, what was missing was a voice. In the 60-year history of Eurovision, there was not a single instrumental song entry. 

I looked at tools to synthesize text and layer song vocals on top of the music. In theory, this is simple: tell the AI – like Boomy or Melobytes – what you want (male or female voice, rock or more opera, etc.), and plug in the lyrics Chat-GPT had created for us.

The reality is a genuine nightmare. Some songs were amusing (a random, disconnected fever dream) or pretty scary (with distorted voices more suited to a horror movie). But most were just bizarre. All of them were utterly disappointing. 

Here’s one example created on Boomy.

It became clear that AI that attempts to generate music and vocals in one go simply is not intelligent enough to do this well.  

I then set out to create music separately from the vocals, and let AI mix both together. To do this, I attempted to turn our Chat-GPT-generated lyrics into vocals by providing instructions like key and tempo. But even my very best attempts were poor and disappointing, lacking rhythm, emotion, or any resemblance to a human singer.  

It was time for another change of approach. There are a number of really clever tools that turn written text into spoken words, with accents and dialects if desired – Speechify and NaturalReader are good examples that can do this effectively. Some tools can squeeze words into rap patterns by dragging out syllables and adding pauses between words. That’s pretty clever and intelligent – and it was a (kind of) rap song that won last year’s contest – so why not create one? Here’s what the Uberduck tool managed.

But, Eurovision isn’t a rap battle, so my rap songs needed something else to embody the poptastic standards of the competition. And here I ran into a familiar problem. I could not find a way to auto-create structure and build-up, as well as add melodic sections. 

AI and human intervention

At this point, I had one realization. If I had better skills and knowledge of how to use the tools, maybe my attempts would have been more successful. But unless I’m missing something completely, there is currently no way to instruct an AI to create a song that sounds unique and human. A few individual elements, maybe. But the sum of it all? Absolutely not. 

Listening to entries to a recent AI song contest, there are some really great examples. But none of the songs were created with 100% AI. Every single one had human intervention in terms of editing, composing, mixing, or singing. And a lot of the tools that were used, like Jukebox AI, require a decent amount of coding experience. Some tools are very good but they’re not for beginners. 

You really need to have a good understanding of music composition, recording and mixing to get a decent outcome. Plus, a very good understanding of how music works and a generous amount of creativity. So, the assumption that AI can easily replace humans to compose and write songs is wrong. And one more thing: the best AI songs I could find, lack soul and emotion. As good as they may be technically, they just feel “off”. 

Am I disappointed? 100% not. It comes with a sense of relief that computers cannot do what we can do as humans, and that it does require a huge amount of our input and expertise to achieve the desired result. Feeling the emotions that a song evokes makes us human, and the ability to create songs with emotion is, at least for now, firmly in human control. 

Will AI take our jobs, as some media reports suggest? There are definitely areas where AI helps us increase efficiencies in our roles. But when it comes to adding real value and creative talent to what we do, we’re still unbeatable.  

Composing music is one for us humans only, but how does AI fare when it comes to creating the visuals for the band? Let’s find out in the next stage of our Eurovision journey.