Music as therapy (or… therapy stand-in) is a tale as old as time. I’m convinced the first cave-teenager to suffer a break-up took to banging rocks together in a rhythmically satisfying way, rather than talking about it.
I could absolutely have been that cave-teenager. Finding music that resonates with the soul, or at least the big feelings of the moment, is cathartic in a way almost nothing else can be. Even if it does take a bit of mental gymnastics to make every line fit your situation sometimes.
Enter Suno. You may’ve heard of it — but if not, it’s an AI that can create songs.
I’d heard of it before. But honestly? I had simply assumed it’d be little more than a toy. Somewhere along the line I got the idea you had to provide your own vocals, and the AI just acted like a souped-up auto-tune.
You can use it like that, sure. But you can also simply write your own lyrics and have it go at it. Heck — if you’d rather, you don’t even need to do that. Just describe what you want, like any other LLM, and it’ll spit out your response.
Only this time it comes with a soundtrack.
Shadows and Neon is one such created in this way. I could not have expended any less effort on it if I tried. I can’t even recall what my prompt was, but I know I didn’t even bother prompting for specific styles. I just wasn’t invested yet.
But it made me sit up and pay attention because I couldn’t hand on heart tell you I’d definitely have identified it as AI-generated if I hadn’t just clicked the ‘Create’ button myself.
Thus Starts my Crash Course in Song Construction
As much as I like music, I couldn’t have told you a few days ago what a ‘prechorus’ was (beyond the obvious), or what precisely a ‘bridge’ is meant to do.
Part of this learning was just feeding in basic prompts and seeing what came out, but I also took a bit of a stab at seeing if I knew enough to take inspiration from other, much better, songs and seeing if Suno could get me anywhere close.
I was even reading song and album reviews for crying out loud. It was a bit like learning a new language adjacent to, but not quite the same as the one you’re native to.
What Never Relents is one such attempt, where I was looking to see if I could craft a style prompt to hit a similar feel to Specter by Bad Omens. It does tickle me that the generated lyrics here used ‘Specter’ without specifically being prompted to or given the lyric directly.
While not an unqualified success (particularly not on the vocal stylings), I was, once again, somewhat impressed. And I hadn’t even realised that you could enfold section-specific direction to vocals and music in the lyrics as well yet!1
Toy to Therapist
At some point, a switch flicked. Don’t ask me how or why — I couldn’t tell you.
What I do know is that I realised I’d still been treating Suno as a toy, when it was clearly more than that. So… I did something I haven’t done in many years.2
I wrote something like poetry — only this time, I had a way to give it a soundtrack.
I think we call those songs.
And because she’d been on my mind a lot lately, I wrote one about my mother. Let me tell you — I was not ready to hear the words sung back at me. Jesus H. Christ. I’m uncertain what context to give. Perhaps none, and leave it to what I’ve already said in the past. Songs are meant to speak for themselves anyway, right?
It hit me harder than I expected. Harder than I was ready for. And yet, I kept working on it. It took a fair bit of finessing of lyrics and style guides to get where I wanted it — and Suno really does not like being told what to do if you introduce the concept of a second voice — but I ultimately found my way to this:
While it’s still a sombre piece overall, the more I tweaked, the more small notes of light started slipping in. Verse 3, for example, wasn’t in the first draft at all — it only emerged as I kept writing.
What Next?
Well, for starters, I think I finally get why some albums feel like they’re circling one idea over and over. Apparently, when something matters, you don’t just… wring it out in a single track.
So of course I had to do another. I had more work through here.
And just like that, I think I’m on the path to just creating a whole (virtual) album. Whodathunk? Certainly not me.
I’m not blind to the issues such AI technologies pose, whichever side of the fence you sit on. I can’t say I know how ethically the models have been trained (but… I could guess). Nor do I relish a future where the world is drowning in ‘AI Slop’ in the music field as well.
I think the transition of considering Suno from Toy to Therapist is perhaps telling of my perspective.
But to spell it out; I don’t think anything I make with it should soar to the top of the charts, or replace the work of genuinely talented artists.
But as something personal — a way to process, to experiment, to speak feelings I’d struggle to put into words otherwise?
It’s powerful stuff.
The first cave-teenager banged rocks together to make sense of big feelings. I guess I’m doing the same, only now the rocks have been swapped for an algorithm and a pair of headphones.
Bonus Track
Second (made, but maybe not in final placement) track from this hypothetical album.
Honouring Mum’s request for how she wished to be remembered — but without burying the truth for those left behind, either.
It was a trickier one to land, asking Suno to balance minor and major sections to hit the right overall tone. Last night I was just about ready to come to the conclusion Suno simply couldn’t do it. After changing my approach to the in-lyric style notes and the main style guide, here we go!
The chorus of Remember Me makes me think that this particular AI must have been fed a lot of Nickelback. ;-)
Seriously though, while I’m standing very firmly on the ‘can we please stop this AI nonsense already’ side of the fence for all the obvious reasons as well as a couple personal ones, I’m happy for you. And I can very much relate to how hard the song about your mom must have hit you.
Music is basically the best therapy there is, that much I’m sure of, so I certailny won’t begrudge anyone using the technology like this. As least it does something very good and useful for a change.
Haha! Dangit. I’m not going to be able to unhear the Nickleback in it now. Thanks for that! xD
If I was to describe my feelings on AI at the moment, I think most succinctly it would be, ‘It’s complicated’.
For better or worse, I don’t think this ‘AI nonsense’ as you call it is going to go anywhere. Well. Depending on how heavily you’re pairing the ‘AI’ with the ‘nonsense’. I think things are going to get worse on a number of front before things get better again — but I guess I do have hope that the end result will be ‘truly’ better than where we are today without it.
Either way, totally with you on Music being the best therapy. Like I said, going into the experiment with Suno, I didn’t expect to find anything more than a toy, and yet… here we are.
I don’t know if you’ve noticed me mentioning it in various posts but Suno has largely replaced gaming for me for most of this year. It’s incredibly addictive, in just the way MMORPGs were twenty years ago, except you have something of real, lasting worth to show for all that time spent.
I have more than a hundred finished songs, all of them from uploads of either actual songs I wrote and recorded in the seventies and eighties or ones I’ve been writing recently, all dervied from fiction I wrote in the 90s. 100% of the lyrics and vocal melodies in all of them are mine, along with the rhythm guitar in the olde recordings. The new ones I either just sing unaccompanied or when my aging voice isn’t up to it, whistle. I’ve always been a pretty good whistler.
Suno is really odd to work with. Sometimes the very first result is almost exactly right, sometimes it takes me dozens and dozens of iterations and a couple of days solid graft to get what I’m after. The end product is frequently amazingly good, though. I’m trying to figure out what to do with them next. And the stories they were created from, too…
I have seen your Suno references! And in fact; I’m somewhat certain I’m only even aware of Suno through you, and that it likely planted some seed sometime ago that finally combined into deciding to experiment with it.
I had also, separately, seen you comment on gaming less recently, but I didn’t explicitly recall the Suno tie in there. Now that you’ve mentioned it though, I’m caught between ‘Oh yeah, I think I did read that’ and ‘Is my brain making up memories to fit the narrative now?’ xD
In any case, dang! Impressive.
I had pretty poor results with trying to use other audio sources as an influence (mostly just tried using existing generated tracks using the ‘Persona’ function, but it seemed to hard cap the length of generated songs shorter than I would’ve liked, even with the ‘Audio Influence’ set to very loose settings.
Beyond that, I have narry a musical bone in my body, so could never hope to write the music itself. I am quite happy with the catharthis that comes from working on the lyrics though, then the style guide as something of a director/producer role, and seeing it to come to some form of life out the other side.
I’m finding it a bit difficult to work with though on some very specific things, even with the in-lyrics [style tagging], specifically when I want to mix vocal styles throughout, e.g., a pure-clean verse 1 or bridge, with grit/compressed voice elsewhere, and perhaps as a very rare ‘accent’ word or line, a full growl or scream.
As soon as you introduce guides toward those elements, they seem to just pop-up anywhere Suno likes excessively, unless what you’ve asked for is a pure/falsetto section, then 9/10 it’ll just ignore it completely.
Still- even with these limitations, I’m impressed, and even though just started with 4.5+ Beta, I can’t wait for future models. xD
I should definitely complain about things more often. Not much longer after I posted that comment about not being able to get the mix of voices styles I wanted in a song, Suno did it. Switched from asking for Falsetto to Clean Head Voice, which, while not exactly the same, Suno seemed to understand better.
Even gave me the growly near scream in just one spot (bridge), and applied grit only where I asked for it. So, colour me surprised, and yet further impressed.
Of course; it’s still a bit random, I generated a few more after the one I liked and published, and while they were all definitely better none of them quite hit again (in the few gens I tried) exactly what this one did.
The Shadow Had Its Way.
I find Suno both incredibly smart and astoundingly stupid at one and the same time. Maybe ten per cent of the time the very first result is almost exactly what I asked for and no amount of refining can really improve on it much. Mostly, though, it takes a lot of iterations to get it to do what I tell it to do. I’ve had all kinds of theories about the best way to prompt and I’ve read lots of other people’s ideas on it but the problem is you find a way to phrase things that seems to work really well and then suddenly it doesn’t any more. Very frustrating.
I do also think that adherence to style instructions has gotten noticeably worse of late. I stoped using 4.5+ and went back to 4.5 because so many of the + results were just plain bad or wrong or glitched. 4.5 is a lot more reliable. I’m also fairly convinced that the AI used to be able to impose a given style on an uploaded recording without too much trouble but now it frequently seems to ignore any style instructions at all and give me the style it thinks is appropriate to the source material. If I iupload a slow, mournful folkish melody and ask it to give me an upbeat, fast-paced pop song it just doesn’t even attempt to do what it’s asked. It’s like it thinks it knows best!
Still, that’s all part of the fun. And the final results, even if they take some time to get to, are extremely impressive.