The metric I was aiming for was not quality, though that’s the most common comment that I hear, including in the YCombinator news website that’s talking about this article. If that’s the case, then perhaps an overfit model where every output sounds the same and mirrors the input pop music data would satisfy those evaluations. The simple model that I built was designed instead to address the 2 definitions of pop music that I thought were integral to its identity. As such, I’m more concerned about if those definitions themselves are correct, and if my model meets those expectations. It’s great that many others have pushed the conversation to those directions, to much of my enlightenment.
Now for you own music, it sounds great. Would love to learn more about your model and other demos coming from the same model.
Cheers.
