So tonight, I'm on yet another obsessive coding bender, and this time, I'm diving deep into the fascinating world of monster datasets related to AI-generated music.
I'm gearing up to play with the massive Suno corpus – nearly 700,000 tracks of pure noisy AI-generated goodness.
Now, working with such a huge dataset comes with some data quality challenges. The tracks are noisy, making it a pain in the butt to analyze.
But it's not impossible.
Taking inspiration from Udio's training method, the first step is to clean up the data before splitting it out.
RipXDAW will do the basics here.
Once the tracks are noise-reduced, we can extract sound charts for later analysis. And if the stars align, we might not even need to do the split at all. I'm not 100% sure how that works yet.
But let's assume we still need to split the tracks and organize them into neat little folders and procedurally parse them out as notation.
It's gonna be weird notation (it always is) – we'll clean it up later.
Right now, we're mainly interested in the gradient and the larger trend, not the diffuse static.
Here's the fun part:
I'll compare our AI-generated sheet music against massive datasets of human-created music. Three of them. One control and a chronal reference so we understand the big picture over the last century or so.
Then we eliminate generic progressions that'll mess up the output, breaking it all down into pretty little pieces, we can run the comparison and identify the unique characteristics of AI-generated music and the commonalities between multiple unrelated pieces of traditionally composed music at scale.
Next is a comparison step with the human made datasets to eliminate biases.
There are so many burning questions to explore:
What patterns show up in AI music that don't appear in "real" music?
How unique are Suno 2 and Suno 3 tracks?
Which features are common, and which ones are rare across these sets?
And most importantly... are my songs the weirdest AI tracks?
Let the data lead
Then we get serious, and look for stuff that's either going to be vindicating or damning for the category.
The potential overfitting problem in AI music and how it compares to human music. The extent and similarity of overfitting, if it exists. But, fairly running the same type of analysis over the other two datasets in the interest of even handed unbiased analysis.
I think the only ethical way to handle a discussion on ethics in AI, is to first evaluate the data, and see where it takes us. Right now, we can argue til we're blue in the face in favor or against either side of the argument, but at the end of the day, there are a lot of unknowns here.
We shouldn't be having emotional discussions about this.
We need hard data science.
And then we'll figure out what we believe from there.