Suno is good at generating music, but its voices sound metallic with a dash of high-frequency noise. Which ruins it for me. It's almost there though, I think they will fix it in the next version.
I suspect we will get ML based "upsamplers" / enchancers / artifact removers, similar as we already have for images. That can be used to automatically post process. If they do not already exist?
None of these are impressive in the least. Anything I have heard from Udio is basically trash. It is the AI art equivalent of synthetic cats and pretty face shots. Who cares.
What is ultimately going to be undefeated is training your own model.
I just don't see how any of these examples from Udio or ElevenLabs could be considered good. To me they sound just bland enough that I may be fooled by them if I weren't paying close attention. If I do pay attention it sounds like a song accidentally came about. It sounds like music might in a dream. Eerie.
I think if I gave you 20 pieces of music, half of them made by AI, you wouldn't be able to tell which ones are made by AI much better than guessing randomly.
I am certain I could, in fact I'd bet a lot on it. I'd have a hard time distinguishing AI generated images at this point, but earlier on there were major indicators. This tech is still primitive. The lyrics are dead giveaways, the tracks all sound blended together. A 5 second snippet of a song seems coherent, but the songs never go anywhere. If you listen to a lot of music, and have a decent pair of headphones, it's immediately obvious. Someone who makes music would be able to identify the specific flaws better than me.
Udio, so far, is undefeated.
And ElevenLabs' music demos were very very impressive, but it's still not released.