I trained an audio model that can generate infinite one-shots for music production and turn text prompts into fully playable synths. I'm not only releasing the model but I've also released a video on exactly how I did it (and the inferencing pipeline to let others make text…

(hopefully this is okay to here - it seems like audio models and image / video modeals is allowed but yeah this is a bit different)

So I've been doing independent audio research for a while now. The ultimate dream of this work was actually getting an AI to respond not only to instruments but also timbre itself as separate controllable things.

Think a Grand Piano can sound both Warm / Gritty but also Cold / Sparkly. Its still a piano though.

This level of control wasn't found in any models out there - so I decided to sit down and train my own.

Getting consistent timbre-locked keybeds that actually LOCKS across multiple diffusion calls was hard af but I did it.

I documented the full journey here for those who want to learn a bit or be entertained.

https://youtu.be/x0KnmzH8Mmk

There is also a longer walkthrough if you just want to see the keybeds in action.

https://x.com/RoyalCities/status/2097733712293109842?s=20

No-talk / Showcase only Demo

https://x.com/RoyalCities/status/2097733715543609445?s=20

any finally the huggingface page

https://huggingface.co/RoyalCities/Foundation-1

I've also provided full write ups on the inferencing pipeline associated with the interface so this should allow basically anyone else to go and vibe code their own text to synths if they wanted :)

https://github.com/RoyalCities/RC-stable-audio-tools/