I've spent the last two months tackling fully playable AI-generated instruments. Thought some of you here might get a kick out of the progress so far. The plan is to release everything for free and open source. :)
SUMMARYThe system uses diffusion to keep a coherent sound profile across the keyboard instead of simple pitch shifting. The creator plans to release the tools free and open source, along with a longer technical video covering the training strategy, dataset changes, and engineering decisions.
This video is from my social media where people follow my work but I felt like people may get a kick out of what goes into the challenges here.
Originally I was just planning to upgrade my recently released sample generator to do one-shot generation, but that somehow turned into a two-month rabbit hole trying to generate entire playable instruments instead.
The result is a text-to-keybed workflow where a single prompt generates a playable instrument that can be exported to multiple sampler formats and used in any DAW.
Everything is generated through diffusion - the goal was getting a coherent sound profile across the entire keyboard rather than relying on simple pitch shifting.
I'll also be putting together a much longer technical video covering the training strategy, dataset changes, and engineering decisions for anyone interested in reproducing the approach or for people who finds this stuff neat.
As with all my previous releases, the plan is for everything to be free and open source.