Most AI DAW-related plugins are tools in the same way that Suno is, except the models are smaller. In fact, modern DAWs are covered in AI, if you're aware of it or not.
Even consumer AI, though. Udio was better with control settings than Suno was for a long time. Maybe even now. Suno's got more DAW-like tools. Udio felt more like a Gradio console, which I liked, actually.
Udio was unquestionably diffusion based, meaning that it made music by parsing the prompt conditioning into a latent with the language parser. In this kind of model, your prompt is sort of like a hypothetical, where you're asking "what if these conditions were all true" and accessing that thing in the space of what could be.
The nice thing about being able to run musical diffusion Udio style, was that you had all those settings. Positive and negative conditioning, denoise, and semantic strictness or "cfg guidance" were all present. Udio had their own name for it. I can only go off what I know about diffusion models more generally, and the behavior of the system.
Suno started off as a small (I think) auto-regressive model. You can train an autoregressive to do diffusion, but the way the system behaves is different. The things that slide, the things that melt. How it sounds in a failure state (something I'm surprised none of yall in traditional music seem to have figured out yet).
One tell is the hallucination state.
Music models aren't supposed to be able to speak English. Not generally. When Udio or StableAudio hallucinates lyrics, they're gibberish. They only sound like English words insofar as they were derived from English words. They’ll represent form, not the function of language.
But when I first started intentionally making Suno hallucinate, was that the words... actually made sense. Now follow me here, because this is where it gets interesting.
The better the sound quality got on Suno, the more vivid the hallucinations got. Today, when the system hallucinates, if you can make it do it, you get basically an entire song that sounds like it was hastily written by an intern who learned how to use words by listening to a million pop songs.
On this system, what that means is that something in the toolchain didn't do what it was supposed to do. Maybe the moderator caught the lyrics as a copyright violation, and the lyric sheet went in blank. Or, maybe you asked for a vocally dominated musical category, and told it to do an instrumental.
*whistles gingerly*
Not like I would be intentionally mucking with Suno or anything. But yeah, that's how you can tell. Diffusion models that run on the raw diffusion process are pattern matchers, not reasoners. Autoregressives will reason about things.
One perk to this that Suno has (or had), was wild generalization. It would let you kindle into existence divergent genre combinations that don't or in some cases, can't exist any other ways. You could use the system at the right wildness setting to get some pretty fucking exciting musical sounds that can't exist in nature. My favorite of these was the Harmonifiddle.
Try to imagine that what might sound like.
It makes sense if you think about it, because they're in the same register.
But it sounds so foreign. It has this almost mystical quality to it, because you know what you're hearing is totally fucking impossible. I chose to leave SoundCloud, so that recording is no longer available there.
My ongoing beef with Suno has been that they've been prioritizing the wrong thing with their product. I liked Suno to begin with, because it had this magical quality to it. It was fun to experiment with. Prompt weirdly, remix noise, make avant-garde weirdo things nobody's ever imagined before?
That's why I did it.
Anyway, today what I do is write code, and build out my own musical pipelines. It's the more reliable way to get the experimental tool sets I enjoy working with them.