So, I’m not very good at creating synth patches, especially for synths Ive designed and built.
I thought maybe AI would be better at it, after all it understands the classic synths and architectures, it knows the Teensy Audio Library and its modules, what could go wrong.
I created a list of all the parameters, there possible ranges, switch positions etc, timings for envelopes, cutoff ranges, even 7 example patches that I had created and what sort of sound they made.
The results were dissappointing to say the least, I spent a few days trying to teach ChatGPT and also recorded exmaples of the sounds I had created, all to no avail, it can happily spit out patch files, but the sounds they make are truly awful.
So 3 days of my life wasted, I could have probably created 20-30 good patches in that time.
I am not surprised. My guess is that there is lack of data that the model could have been trained for that sort of task. I mean, I have probably read dozens of books and visited thousands of websites about synthesizers, but I never came across any information/explanation about what sound is produced by which parameters.
It seemed pretty well versed on the patch types for various synthesisers, it talked about the types of sounds it wanted to try and create such as oberheim brass and big polys etc.
I recently tried some SVG file manipulation for a pcb project with chatgpt. it made bold claims about what it did and could do. provided mulitple options for further improvements…
I agree the results firom AI solutions can be shockingly bad, I asked chatgpt to review my analogue mux code to make it more stable and it made it 10 times worse.
I got AI to generate a schematic once for fun and it spit out something with an op amp with two non-inverting inputs. Probably not the best for anything complicated.
LLMs thrive when there is plenty of training data thrown into them. Think of million, billions, lines of code from github and a multitude other sources. Millions, billions, of annotated images. Equations, and encyclopedic knowledge from wikipedia and online books. All of them with plenty of contextual content. Opinions about this and that culled from online forum discussions. Etc.
Instructions about what parameter needs to be what value for whatever type of sound is not something that we have in abudance online or offline. When was the last time that you noted the knob positions of a patch that you made and shared them online? Is there a place where people share such information extensively? Do you actually know of a place where you can download data in a non-proprietary, easily understood by a language model, format about the preset patches of well-known synthesizers?
I am not surprised that we are lacking such data either. Most people cannot be bothered to programme sounds and just use whatever factory presets are available. Even people like Vangelis could not be bothered. The CS80 presets are all over his albums!
