Suno Experiments
Before I started an album, I spent a while just poking at Suno to see how good AI music generation had actually gotten.
I found Suno last year and impulse-bought a one-year pro license, the way I impulse-buy VSTs, DAWs, or projection-mapping software: to see if the workflows have matured enough for an amateur. Can I get from an idea to something polished without spending two months learning the software? I keep hoping DAWs like Ableton or Bitwig will evolve the way Blender did, and usually end up frustrated.
Like most people, I started making goofy songs about my cats with sparse prompts. Fun for a while, then boring. The spot-editing controls weren’t much use either: the granularity for a small change, or pushing a variation in a direction, is still limited.
The turning point was feeding Suno my own material. In the mid-2010s I recorded a few songs with friends; I uploaded those and was amazed at how much better the output was. It did the work I used to find tedious, the mixing and production.
Repeatedly “covering” one of them, Time Machine, produced a pile of note-faithful remixes and genre adaptations that sound incredible. (The lyrics are based on a public domain poem whose author I’ve had no luck tracking down. If you know who wrote it, please reach out.) Here’s the original (forgive the drum machine, we didn’t know any competent drummers at the time):
The first good cover; notice how closely it tracks the original:
A dubstep version:
A donkstep version (a Russian EDM/hardbass crossover, lol):
And my favorite, a tech-house cover sung in my own voice:
The workspace problem
The fun part scales better than the interface. One song turns into covers, then covers of covers, then voice-model passes and genre detours. My current workaround is one workspace per album track, plus a “Collect for Final” workspace and another for errata and old versions. That is organization in the same sense that a junk drawer labeled “cables” is organization.
The numbers get stupid quickly. The workspaces for Fyodorov’s Dream and Census of One contain 90 and 51 songs. Time Machine has 49. Inside a workspace, uploads, covers, voice passes, and generated tracks sit in one long list. Many still have names like “04 - new,” so the title does not tell me which version descended from which. Search and filters help me find a row; they don’t restore the family tree. Even deleting rejects mostly moves the problem into another enormous list in Trash.
Remembering which generation had the right vocal, what it came from, and which branch I meant to continue is manual archaeology.
There are entertaining detours in that pile. I have been making private hardbass covers by giving the model Cyrillic transliterations, which makes it pronounce the words with a Russian accent. The same text sometimes gets different filter treatment in Cyrillic. This is messing around for my own amusement, not a release strategy. Covers of copyrighted work stay private and I don’t publish them.
The useful experiments are mixed into all of this with very little structure. That frustration is part of why I started moving the durable work outside Suno: lyrics, production notes, motifs, and lineage in files I can track myself.
I started noticing patterns in what covers preserve and what they reinvent, and got thinking about using intermediate tools to scaffold a song, or a whole album, from scratch, then produce it with Suno plus other software. So I decided to tell a story: Wavefunction Collapse.