New Site Uses Artificial Intelligence To Create Yamaha DX7 Synth Patches So You Don’t Have To

The Yamaha DX7 is one of the most popular synthesizers ever made, because it offered an unprecedented combination of polyphony, playability and pricing when it was introduced.

It also earned a reputation as one of the least user-friendly synths to create patches on, because the DX7 used an approach to synthesis that was new to keyboardists at the time it was introduced and editing was slowed by a menu-heavy interface.

While many synthesists have taken the time to learn DX style FM programming in the subsequent years, a new website promises to make creating custom patch banks faster than ever before.

The site, This DX7 Cartridge Does Not Exist, is fount of DX patches. It uses an artificial intelligence model, trained on a collection of hand-crafted DX7 patches, to create synth patches on the fly.

Here’s what the developer has to say about it:

“The Yamaha DX7 has a total of 155 parameters, the AI uses a machine learning model known as a VAE that reduces these 155 parameters down to just 8.

The model can then choose these 8 numbers randomly and map them back to the full 155 parameters of the DX7, it has learned to do this by training on a large selection of presets compiled by Bobby Blue

Each cartridge is generated in real time by sampling 32 sets of the 8 VAE parameters and compiling them together in real time. The model generates 99.9% unique voices and so it is highly unlikely that any two cartridges created on this site will be the same.”

The site generates a MIDI .syx that you can download and send to your DX7. It’s also compatible with Dexed for Linux, OS X and Windows and Rockrelay Synth FM for Android.


18 thoughts on “New Site Uses Artificial Intelligence To Create Yamaha DX7 Synth Patches So You Don’t Have To

  1. Good to see it works with Dexed. Am I right in assuming the same for Arturia’s DX7? My current favourite for this sort of thing. It would be interesting to see what this comes up with. Although, as it is, I have literally tens of thousands of DX7 patches collected from the web already, plus a few hundred I’ve programmed. Although, like modular, FM is on of those twisty turny modes of synthesis, that often throws up some unexpected curve balls, so to speak, giving stuff you hadn’t thought of. So seeing what AI makes of it is interesting in and of itself, especially if you use any AI generated patches as templates to further work on. Sounds like an intriguing project.

  2. Cute. Tried one cart in Dexed. Wonder which proportion of patches people would deem usable. Didn’t find one which really jumped at me and several were completely unusable or too similar to one another (as short bursts of almost-white noise).
    There’s a lot of hope and hype in AI. There’s probably a lot more to do in collaboration between AI and human curators. Although, I’m not always convinced that AI is much better than a bit of controlled random.

    1. Yeah they are mostly unusable sounds. I like that it has AI’d the patch names too. I’m guessing they fed their algorithm all the original DX patches plus all the Dexed carts available online. Fairly obvious it was going to just mash them all up. Would be more interesting they had fed the AI custom carts containing only bass sounds, pad sounds, piano sounds, percussion sounds etc.. So instead of just randomly mashing up every type of DX patch, it was mashing up DX patches of a similar style. Perhaps then the output wouldn’t be so random

        1. I would like to give this comment 100 points but there are only 6 available operators in various yet fixed configurations so I can I can only give it a total 32 well-deserved points. Now if we add operator feedback…

  3. Haha, no one likes to program FM, they finally got a computer to do it for you. I do like programming it though and learning more about synthesis. They didn’t really let anyone know the details of that they should study sound and there is a whole other book to some books on the subject.

  4. It would be very interesting to know what the 8 parameters sound like when twiddled. That would be more interesting than just generating random combinations.

  5. Fun idea. A properly trained model should produce a million variations of brass, bell, and electric piano patches.

  6. Would love the ability to set the ‘learn’ set to some fixed set of patches instead of the universe of Sir Bobby’s masterfully collected patches. Then you could make a set of 32+ ‘bass’ sounds and have the AI learn from that. Or a set of electric piano sounds. or tine sounds… or…

  7. “New Site Uses Artificial Intelligence To Create Yamaha DX7 Synth Patches when you have no Intelligence of your own…”

    cool. maybe ill skip that tho. Cause, like… you know….i have my own intelligence and stuff….

Leave a Reply