It seems to struggle to create music with a strong identity. It is great if you want to make a poor imitation of top 40 hits. But the thing about top 40 type music is that the best music is already in the Top 40. It remains to be seen if there is as strong a demand for a music chart filled with slop as there is demand for a music chart filled with pop tunes by celebrities.

I don't think audio files are the right output for deep learning music models. It'd be more useful to pro musicians to describe some parameters for synths, or describe a MIDI baseline, or describe tunings for a plugin and then have the model generate these, which can then be tweaked similar to how we now code with LLMs. But generating muddy, poorly mixed WAVs with purple prose lyrics is only an interesting deep learning demo at this point, not an advancement in music itself.

> It is great if you want to make a poor imitation of top 40 hits.

generation models in a nutshell

Since we've already evaluated - let's exaggerate - that "most Top 40 songs are slop", maybe lyrics are a big factor in creating identity? I mean, it's true for books, right? I could easily imagine an AI-generated Top 40 song that people would still describe as having a unique identity.

I'm not super into the topic, but let me give you two niche examples that are definitely not Top 40 material, yet are considered to have a strong identity within their communities.

I guess one of the reasons the game Yasuke Simulator has like 10x more sales (don't pin me down on that) on Steam than the actual game Assassin's Creed: Shadows is its very catchy soundtrack, with lyrics that are funny and strongly aligned with the content. [0]

Another example, not focused on lyrics and from a completely different niche genre, is this jazzy death metal song that was particularly well received, not only because of the intentionally hallucinatory video. One could even argue that the hallucination is perceived as a feature, not a bug. So why shouldn't the same be true for audio? [1]

[0]: https://www.youtube.com/watch?v=Hkh38jhILec

[1]: https://www.youtube.com/watch?v=OzXafFUekl0

I actually built a very crude python app that could generate basic melodies with MIDI. It can serve as a fun starting point if you want to remix common songs. But it wasn't a very fun project to set up. It's like this technology is out there but nobody really wants to develop it.

If I had to guess there are already a handful of fake record labels generating at tons of AI slop to just post on Spotify. Even if each song only gets something like two or three views over time they can still generate a modest amount of revenue. Oh wait Spotify has been caught doing that themselves

Some of the artists of the top 40 use some form of auto-tune to generate pleasing music. Can that be considered a slop-y?