Music and AI
by David Barrat

From computers playing chess, the software for which was widely commercially available in the 1980s, it's been quite a long and slow process for artificial intelligence to develop further uses, but it's here with a vengeance today, threatening to take over from human beings in multiple fields, putting people out of work in the process, and, perhaps worse, threatening to confuse as all as to whether what we are seeing and hearing with our own eyes and ears is real or fake.

​

It hasn't all gone smoothly, of course.  Lawyers using AI for legal research purposes have found themselves sanctioned for providing judges with fake citations to case law which AI had simply invented, for reasons best known to itself, if it really "knows" anything at all.  When the organisers of the Willy's Chocolate Experience in Glasgow decided to save money by using AI to write a nonsensical script for the experience, it all turned out to be a humiliating disaster.  And there have been many other problems with the use of artificial intelligence in so many different fields.​

​

AI in music may not be entirely perfect and there are quite a number of legal issues which appear to be unresolved but, in terms of how it can produce incredibly high quality songs at a fast speed, it is quite remarkable. 

​

After trying Udio but not really liking it, I first started uploading my music to Suno in August 2024.  At that time, users were limited to uploading tracks lasting no more than one minute, and Suno would only extend tracks not replicate them.  In other words, it would produce something melodically similar to what had been uploaded but not the same.  The "cover" feature on Suno was introduced a few months later, by which time it was possible to upload tracks of up to two minutes in length.

 

So the Bubblegumelicious tracks have, in general, been created in three ways. 

 

Sometimes, in the early days, I would only upload a portion of the instrumental of a track (occasionally just the middle 8 section, if I happened to like that section), together with the lyrics, and see what melody Suno would come up with.  Often, but not always, it would come up with what might be considered an improvement on the original but always based on the original instrumental backing track. On a few occasions, I forgot to click the "extend" button so that Suno would generate a new version of a song based on the lyrics only.  Interestingly, every time that happened, I found the result uninspiring and dull.  True, I could have continued to generate alternative versions with those lyrics, and perhaps it would have come up with something better, but the fact that the first attempts weren't much good, at least in my opinion, led me to the conclusion that what is input into Suno is a critical factor in achieving a good output.

​

The second method was putting in a version of the track containing vocals and requesting Suno to extend the track.  Again, the results were invariably similar but not exactly the same as the original melody.

​

The third way of doing it, which became the primary method once the "cover" feature was introduced, was to give Suno two minutes of the track, with vocals, and request a cover version, almost invariably in the pop style, or a variation of pop.  Usually, the closer Suno's output came to the original version was the best result. But it was always interesting to hear new arrangements of my songs in various styles.

​

In most cases, uploading two minutes of a track meant that the middle 8 section, which usually appears in the third minute of a song, couldn't be provided to Suno.  Where I thought a middle 8 was particularly good, I would start the track from the second verse in order to give it that section and leave it to work out the first part of the song on its own based on the second part.  Occasionally, in circumstances where it didn't know my middle 8, Suno would, by chance, produce a middle 8 similar to the original or, perhaps, not so much by chance, because there are certain musical rules which guide what a middle 8 of a song should sound like, but I usually found that Suno's middle 8 sections weren't as good as the originals.  Nothing, however, could be done about that.

​

Most songs I originally extended were subsequently covered, once that feature was introduced, and the results compared to see which was the best version.  Sometimes the Suno melody won the day, sometimes the original melody won, but the differences between the two were usually not so great and the Suno melody was never too far from the original, being based, of course, off some part of the original version of the song.

​

I wanted to set out this brief summary of how the Bubblegumelicious songs were created to try and clarify the role of AI in their current existence.  Every song had already been written and demo'd, and thus existed in some physical form before going near Suno.  No AI lyrics have been used.  Only AI vocals and AI musical arrangements. 

​

The original melodies input into Suno were almost always all mine.  On very rare occasions, the arranger I was working with might adjust the melody of one of my songs but those can be counted on one hand.

​

The sound quality of the songs that Suno has produced has been pretty good although some of the tracks are a bit muddy and there are improvements that could be made in this area.  Suno's v.4 has certainly improved some of the quality of the tracks but they're not absolutely perfect and the vocals often suffer from distortion on held notes, which is less than ideal.  Overall, though, I would say that the results are what one would expect from a top quality recording studio.

​

From my perspective, while being in awe of the speed and quality of output of Suno, what it does is nothing new, in the sense of what I end I up with.  I'm very used to giving a freshly written song to a producer/arranger and having them work their magic to create, with the help of a session singer, a fully arranged and produced demo of that song for me.  On a few occasions as I've mentioned, the producer would change the song, select new chords and melody to give the song a very different feel: a process which has always fascinated me. It's incredible how a new arrangement can transform a song.  Essentially, all that is happening now is the same thing but instead of giving my song to a producer/arranger, I'm giving it to an AI tool which produces a number of different versions with different arrangements and different singers for me to choose the best one.  It's not really any different from what I've been doing all my life, except the process is so much faster (and so much cheaper).  I guess it's kind of like me giving my song to twenty different human producers and them coming back to me very fast with twenty different demos of the song for me to choose my favourite.  It's exactly the same amount of work for me in creating the song in the first place.  

​

The main advantage of using Suno is that it has allowed me to change certain lyrics which I've never been entirely happy with in recorded demos.  Some lines did bug me when listening to the recorded tracks knowing that I had a much better alternative lyric in my head. Re-doing a song outside of Suno, just to to amend one word, or one line, would have been prohibitively expensive and time consuming but it's now so much easier to do this. Theoretically, though, I could have done it at any time.

​

Human or computer, then, doesn't really make any difference to me in terms of the value of the song.  The AI is only "better" than a human in the sense of speed and cost.  Working on Ableton Live, I've used many plug-ins which assist with song writing.  My favourite is Scaler but I tend to use this on songs I've already written.  I'm aware that there are tools which can generate chords and melodies which can then be used as the basis for writing songs.  AI is no more than another songwriting tool, at least in the way I use it.  Asking AI to generate a song out of nothing, or feeding it lyrics generated by AI and then asking it to produce a song, is another matter entirely.  But, as far as I'm concerned all other songs, not written by me, are exactly the same, whether written by humans or by artificial intelligence.  The process isn't important, only the outcome.  If AI generated songs are better and more enjoyable than songs created by humans then so be it, I'll listen to AI generated songs.  But I think we'd all prefer it if the songs we listened to were made by other humans.  You can be sure that this is the case with all Bubblegumelicious songs despite the input of AI.

​

Naturally one hopes that humans will win out and that AI will never be able to capture elements of genius that human beings can incorporate into songs and will be unable to adequately reflect the human experience in an emotional way to touch people in the way that great songs can, but we'll have to see.

​

David Barrat

December 2024​

​

Back to Home page