AI and Lyric-Writing

 

When I generate AI music from lyrics, they are usually in a provisional state throughout the process. 

There was a lyric I was working on recently that contained a factual error, which I caught later on. At that point, the music was where I wanted it to be, but the error was the fly in the ointment. The question becomes, do you keep it as-is and include a "footnote" about the error and other vocal flubs or "mechanical hacks", or do you keep regenerating it until you get a piece of music that works? With a real vocalist, you can ask them to do another take with corrected lyrics.

The lyric at issue was “anonymous artists from Banksy to Pak”. Pak isn't an anonymous artist, but his name is conveniently one syllable and works rhythmically. When I chose a different artist name (Blu), the music that was generated after that point wasn't as good as the original version. I could change "anonymous artists" to "pseudonymous artists" but that's clunky and difficult to sing and has probably never been used in a lyric.

The larger question is whether I want a "casino creativity" in which composing becomes more like playing a slot machine. Eventually, you’ll win, but how long are you willing to play the game? I will admit that there can be a payoff when you like the results.  Kris Kristofferson croaking lyrics about NFT artists, minimalist white cubes, and Vantablack with a Morricone vibe--fantastic...

Regardless of how generative AI is uncreative, perhaps there is still a randomness at play. On one of the iterations, a particular vibrato on the word "machine" sounded like a machine and was an amusing happy accident. But a long "e" when sung with a vibrato will actually sound like that.

There are also some lyrics that simply cannot be sussed by AI, no matter how many times you try. It can generate jazz, but it doesn't understand swing or syncopation, or its version of it is strangely artificial.