Writing a song using AI

Writing a song using AI

I decided I’d try my hand at writing a song with AI. Just so you know, I’m a lyricist that never writes melodies – I know too many great composers. For the exercise, I decided to write one that would fit the brief for a perfume ad. I was looking for a late-sixties vibe, the type of thing that would be incidental or theme music.

Following a recommendation from FutureAILab, I opted to try Suno.

As the site itself points out, Suno was largely developed to create music on original lyrics. Well, that’s handy. I’ve got tons of those.

So I pulled out a text I had started recently, that happens to mention a perfume – although that was purely coincidental.

Prompt me, baby

As we all know by now, the secret to mastering AI is the art of the prompt. For this, I gave it a relatively precise indication of what I was looking for. “A moody instrumental cool-school jazz vibe, with a deep male voiceover.”

After a minute or two, I got two versions. Both are OK, although I preferred this one:

“Something” co-generated with Suno

The result is not bad, but I can see how I would improve it already (see below).

It ticks a lot of boxes for me. It does have a 60’s vibe, when singers were often doubled up on lead. It’s not exactly Francis Lai, but it’s good enough.

But good enough for what? I was using the free version, which is billed strictly as non-commercial. It’s not a commercial pop song, but then the prompt didn’t ask for one either.

Also, I received it as a mix. So if a client or singer says, “Nice, but could you just…” I’d have to record the track. In this case, with such pared-down instrumentation, that would not be a major problem as we know exactly where we would be going. With a pop song with effects, that could be more complex.

The result of writing a song using AI has a certain randomness to it that composers might hate. But as the lyricist, this is no different from what I experience every time I hand over lyrics. I rarely have a precise idea what will become of the words!

Learning from AI

One lesson learned is that the words should ideally be really polished before submitting. When I work with composers, I’m used to a fair degree of interpretation and changes. Parts are doubled or deleted depending on the circumstances. Lines are extended or shortened for the singers. Here, what you give is what you get. I really should have added some variations such as “Something – something about her eyes” to brighten the melody towards the end. Maybe I could introduce some written hesitations to make it more human.

The structure of generative songs depends entirely on the lyrics. So that has to be worked on carefully before submitting. I have since discovered that you can vastly improve the result by using verse/chorus/bridge and outro tags in the lyrics.

I also have to be much more careful about choosing the main genre, which requires me thinking more carefully from the outset. What do I want to achieve?

Lastly, I’ve been listening to a fair few generative music tracks and playing around some more. One thing that has to be done after receiving a track is to get it mastered. That could really lift them out of the AI-sounding domain.

To sum up, I have found a handy assistant for creating song demos from lyrics that never found a home. The songs that are generated could be used out of the box in some situations, such as commercial jingles or generic style playlists.

There is more to writing a song using AI than it first seems. I’d really love it if you could try the tool too and share some experiences (and links) below.

Lyrics by Michael Leahy

Leave a Reply

Your email address will not be published. Required fields are marked *