"Can You Sing It A bit More Like This...?"

Time for the next episode in our thrilling story! So far it's been all set-up and no action. Will today see some new music being made at last?

Maybe, although I see three potential points of contention just in that last sentence alone. 

1. How "new" can music be if it was written and recorded forty years ago? 
2. Is it really "music" if an AI is making it? 
3. And can an AI actually "make" anything?

I'm going to take a bit of a position on this one right away. Having spent many hours with the tools and having listened to the results, I'm going to say yes to all three. This is music. I did make it. 

As for it being new, it's new in the form it exists now, although it also remains time-locked in some essential, existential way. It's sometimes seems to me, as I listen to the new versions of these old songs, that they were written by someone I scarcely recognize. Someone who isn't me any more. It's the good old intentional fallacy brought to light yet again.

I'll get to that later, probably. First, the mechanics. So, in the last two installments, I reported on how I got the old recordings off tape and onto my hard drive and how I decided what software to use to try and turn them into something that sounded like the music I was hearing in my head.

At that point I was under the impression I'd be able to upload the files to one of the AI apps and tinker around with it to get what I wanted. That turned out to not to be the case at all. Not at first. 

Like that's going to help...

The mistake I made was to assume that the "Edit" option in Suno would allow me to do things like specify instrumentation and vocals and have the AI replace my strumming and singing with a facsimile of someone who could do it properly. Then I could ask the AI to add drums and bass and so on until I'd built up a composite version that sounded how I imagined it should.

This is categorically not how the software works. There is an Edit function but it doesn't really do any of that. It lets you extend the song, replace sections of it, crop it, fade it and change the lyrics. It's still the same damn song, though. 

I played around with that for a bit but it was totally useless for my purposes. It was still me singing and playing and it still sounded like a bad recording of a bad singer and a bad guitarist.

I was pretty fed up at that point and on the verge of giving up on the whole idea, at least until the technology advanced some more, when I happened to notice, purely by chance as I was fiddling with the settings and pressing buttons to see what they did, that there was something called a "Cover" function.

I looked it up on the Suno blog and here's what it says:

"This feature lets you take anything from a simple voice memo to a fully-produced track and transform it into an entirely new style, all while preserving the original melody that makes it uniquely yours."

There's a lot more. You can read it all at the link if you're interested. The tl:dr is that the Cover function does exactly what the name suggests: it creates a cover of whatever you feed into it. It's just as though you'd asked a performer or a band to cover a specific song, given them a recording and let them get on with it. 

It's magic, basically. Actual magic. 

This immediately had two effects on me, other than making me punch the air and yell "Yes!", obviously. As I began to play around with the Cover function I realized I could either try to get the AI to produce the closest possible match to how I'd always wanted the songs to sound, if I could ever have gotten a bunch of people to play them the way I wanted - or I could indulge myself with endless variations, hearing my songs being performed in all kinds of deeply inappropriate styles, the way Nouvelle Vague or Postmodern Jukebox have been doing for years.

The Cover feature is still in beta.
Can't wait to see what it's going to be when it's done.

Naturally, I did both. I couldn't stop myself. I still can't. I realize now that doing it the way I have has implications I hadn't considered when I started but it's too late to worry about that now. Also, those implications are probably best left for another post altogether.

For now, let's stick to the mechanics. So, how does this Cover thing work? 

It's simple - until it's not. You upload your source material to Suno, click the menu option for "Cover", specify whatever style or genre you want, along with any other specific instructions such as mood, type of singer etc. Then you hit "Create" and sit back and wait.

Not for long. It takes Suno maybe thirty seconds to spit out two covers. You always get two. Suno does everything in pairs. And even though the instructions are identical, the two versions are often radically different.

They are also often radically different from the instructions and sometimes from anything you could possibly have expected. Here's something you really need to know about Suno: it has a mind of its own.

Suno's wild fantasies can be amusing or bizarre but it wouldn't be much use if it didn't do what it was told most of the time. Better than eighty per cent of the time, I'd say. Then, it sticks fairly closely to whatever you've told it to do.

I'm not going to buy into the whole "Prompt Engineer" malarkey but there are some basics anyone needs to pay attention to. Suno doesn't parse long lists of styles and moods well, for example. If you ask it to produce something that's "Indiepop, Dark Pop, Dreampop, C86, Janglepop" it won't usually try to meld all of them together into a pleasing gestalt; it'll pick one and go with that. 

It also doesn't especially like to try and create unlikely style or genre combos like "Twee Funk Psychobilly" although sometimes it can be persuaded to give it a go. On the other hand, it clearly finds some combinations very comfortable, so it will happily give you "Futurepop Kawaii Bass" or "Dark Ambient Vaporwave" if you ask nicely. 

From the very handy
but by no means exhaustive
Music Genre List

Suno also recognizes or can interpret an absolutely astounding number of genres and moods. I've yet to find anything it can't at least have a recognizable stab at, although the results can be variable in the deep woods of microgenre.

One thing I discovered early on is that it almost always follows clear instructions on vocal gender. If you ask for "Female Vocals", that's what you get. 

Trying to get a specific kind of female vocal is a bit harder. I never really wanted to be the lead singer in any of the bands I was in (Makes it sound like there were dozens of them - there were four, in fact.) even though that's what I always ended up being. I always wanted us to have a female singer but we only ever managed to persuade our girlfriends to sing back-up. 

All the same, I've always imagined most of the songs I've written being sung with a female voice, so that's what I hear in my head. I know what kind of voice it is, too, but getting Suno to sing it the way I want to hear it reminds me altogether too much of how it used to be, when I was on the receiving end of "Can you sing it a bit more like this...?" at pretty much every rehearsal.

I tried using just one adjective and I tried using several. I tried whole descriptive phrases. I was never sure which worked better. A lot of the voices were almost right but just not quite. In the end, I ended up using "world-weary", sometimes in combination with "innocent" or "naive", which generally seemed to give me something close to the personality I was after.

Of course, once I'd got it on one song, I had to try and get it again and again on the rest. There is a function called "Persona" that allows you to save a particular voice and/or style and re-use it for other songs, lending them all a consistency that makes them seem like they were recorded by the same singer or band. That was exactly what I needed, but frustratingly, it can't be used for covers of songs that have been uploaded.

The reason for this would seem to be the makers of Suno having an understandable desire not to get sued out of existence by the megacorps. I mean, they already are but why give them the ammunition? There are a number of safeguards built into the system to try and prevent that happening.

I never had any intention of making the covers of my songs publicly available on Suno, which is just as well because it turns out you can't. Even though you have to tick a box to say you own all the rights before you can upload anything, all uploads and covers thereof are automatically blocked within the app from being lited as "Public". No uploading someone else's song then trying to pass it off as an original after you've had Suno "cover" it.

"it's never going to be exactly
what you originally had in mind."

With similar caution, the Persona function cannot be used on uploaded content or covers made from it. If it could, presumably you'd be able to upload your favorite singer, clone their voice and then apply it to anything. If you want to do that, best talk to Grimes. She's up for it. 

This is one of the limitations of making covers of your own songs using Suno. There are several more, probably the most awkward of which is that uploads can be no more than two minutes long. 

I didn't notice at first because the songs I started with were less than two minutes long to begin with. They were the ones I wrote specifically in the late 'eighties, when I was trying to get a C86/Twee band together. Unsuccessfully, as it turned out. 

I uploaded them and Suno happily gobbled the lot. It was only when I got around to some of the longer songs from an earlier phase that I realized it was cutting them off before they'd finished.

In most cases, this is less of a problem than you might imagine, at least if you're dealing with traditional pop or rock song structure. You can select which two minutes of the track to upload and Suno easily recognizes verses and choruses. (You can specify them but I haven't found it necessary.) 

The AI picks up melody, phrasing and even intonation almost perfectly and it can extrapolate from what it knows, so if your three minute song has a traditional verse-chorus structure and just carries on the same to the end, it doesn't matter to Suno that it only has two-thirds of the song to go on. It'll get the rest right anyway.

The problems start when there's a change of some kind outside the part Suno knows - a middle-eight or a coda or a solo or something. Structurally, Suno works from the lyrics, plus any additional instructions you type in in square brackets, such as [Middle Eight] or [Guitar Solo]. When it hits something that doesn't match the pre-existing pattern, the AI will have a jolly good go at coming up with something appropriate but it's never going to be exactly what you originally had in mind, which means that cover is going to have a little more AI influence than perhaps you wanted.

"poorly-recorded caterwauling
converted from forty-year old audio-tape"

There are ways around this, although it took me a while to work them out. You can, for example, cut and shut parts of your original recording so it comes in under two minutes but has all the separate elements of the structure and melody. Suno will then apply the right melody, rhythm or phrasing to the lyrics that match what it knows. It's very clever like that.

And that brings us to the lyrics, where there's another problem to look out for. When you upload your song to Suno, it does its damnedest to figure out what you're on about and write it all down as accurately as possible. If you articulate clearly and don't use a lot of made-up words, it does a more than fair job of it, too. There were a few uploads where all I had to do was make a handful of minor corrections.

On the other hand, if what you're feeding it is mostly poorly-recorded caterwauling converted from forty-year old audio-tape, Suno takes a wild stab at the few bits it can just about make out and gives up on the rest. For most of the recordings, I had to get out my old lever-arch file with the hand-written lyrics and type the whole lot in.

Which I really enjoyed. It was a lot of fun going back over those old lyrics and figuring out what I'd been trying to say. A bit like reading an old diary, which I guess it what happens if you tend to write from experience.

There were several distinct phases, the best, at least from my perspective four decades on, being the time when I seem to have been mildly depressed and obsessed with weather and the turning of the seasons. Or, put it another way, when every bloody song I wrote was about the rain, one way or another.

And that, I fear, is about as far as we're going to get today. This does go on, doesn't it? And there's quite a lot more to come. In Part Four I promise I'll get to some of the actual songs although I'm not promising there'll be examples you can listen to. I'm in the process of making videos for all of them. I've done five so far. I have about twenty more to do.

But that's another post altogether. 

Notes on AI used in this post.

The header picture and two spot illos. All done at NightCafe using Flux Schnell with default settings, other than Aspect Ratio (Changed to 4:3) and Runtime (Changed to Medium.) 

All prompts were taken directly from the text and are shown in the captions except for the header image, which was generated using the quote in the post title plus the next part of the quote "at pretty much every rehearsal", and the color illustration, which also had the instruction "Band rehearsing" appended to be sure of getting something suitable.

I also added a style instruction, "Line art", after all prompts to be sure of getting drawings rather than photographs.