This is a statistically modeled preset generator for the Arturia MicroFreak shamelessly inspired by NeuralDX7 and thisdx7cartdoesnotexist. It tries its best to create realistic sounding presets while not copying existing ones.
Use the encoders or dropdowns to select what Engine Type you want for your Oscillator. Unsure? Leave it at random and let chance decide. Select a category similarly. Categories are based off the assigned types presets can be set to (pad, lead, keys, bass, etc). Hit Download.
If you have an Arturia MicroFreak, you can use the MIDI Control Center software provided by Arturia to upload the single preset to any project and then to the device. Select the preset on the device and give it a spin! I hope it sounds good.
Sadly, there isn't enough data out there to make a deep learning style preset generator similar to NeuralDX7. The Arturia MicroFreak is relatively young compared to the legendary Yamaha DX7.
However, there is enough to make what is called a Gaussian Copula. Think of these like interconnected distributions based on related variables. This models how various settings in a preset are generally aligned or used often. A good example would be the Bass category. A bass synth sound is typically a very short attack and a short to medium decay/release, sustain being of a variable length. The model captures that behaviour, sees they are related, and randomly picks values within those respective distributions so they appear related similar to how a real person would make a bass synth sound.
There is a bit of extra magic behind the scenes too such as measuring the preset against previously seen preset values to ensure it is unique, as well as trying to pick a preset amongst a set of many that is difficult for a detector model to determine if it's human made or generated.
Curious about the model training? Check out GaussianFreak, the public repo.
The MicroFreak comes with a variety of presets packaged with the MIDI Control Center software as well as free packs offered on Arturia's website. That coupled with community collections of various free presets scattered throughout the web and GitHub.
Absolutely. All of it.
If that turns you off, I understand. I've read the code. I understand the models pretty deeply due to my academic background and over 10+ years of experience leading and building data science and analytics teams.
I've done a ton of notebooks and built a lot of services and reran giant model constructions on terabytes of data on massive clusters with hand written code bugs and all well before LLMs. Why didn't I just do that here with arguably megabytes of data? I didn't want to. Simple enough. I do that enough at work.
Sure. Ping [email protected] with any questions.
Unofficial fan project. Not affiliated with or endorsed by Arturia. MicroFreak is a trademark of Arturia.
Background: crewel valance, England, 17th century. Cooper Hewitt, Smithsonian Design Museum (public domain).