coding projects

unknownpasta

I made this to answer the age old question, what if we could make that one Joy Division album cover out of text?

Well, here is a Python module which does just that. You input a string of text and a series of parameters, and the program outputs a joy-division-esque logo. The program generates a new logo every time, since it is based on iterative random modeling of the original.

The results:

synefind

this was a project a friend and I thought of a few years ago. The basic problem we wanted to solve was this:

• we liked synthesizers, especially those in the music we listened to
• we wanted to emulate voicings on our own synths / digital synths
• finding the exact combination of parameters to make a tone is very hard

I had several ideas for solutions, and while none panned out, it was a great learning experience. The most promising idea was to use some generative density estimation model (GAN, VAE, etc.) to learn a reversible mapping from input parameters to output waveform.

It turns out that it is very hard to effectively train generative models on synthesizer waveforms, even after transforming and processing them endlessly. The models which worked best were not good for the target purpose (backtracking to synth parameters) but they did work decently in a generative capacity.

More importantly, the data processing yielded cool plots:

time-frequency visualization of some synth waveforms

Since first attempting this project I have gained a lot of experience working on ML projects, and I actually think this idea would be viable using a combination of more advanced (and easily convergent) techniques.