I made this in 2013/14 for an undergrad university assignment. It was actually a composition assignment but I somehow ended up writing this. I'm revisiting it in 2023 out of curiosity. The code is basically the same as before but I made it work with Python 3 and modern numpy.
I wrote this out basically because I wanted to remember what this thing did, and now I do. I'm glad it still works and does something relatively interesting. It is still just a script that doesn't take any parameters or anything, so if you want to run it for some reason, go and edit the globals in main.py (lol). I'm unlikely to ever expand this, I just wanted it to be able to run again. There is also iota, which is a related and slightly more developed project, which I might actually improve at some point.
- Originally this was vaguely inspired by chance/indeterminacy based music like John Cage etc.
- That guy and his mates had some pieces where they took a bunch of samples (originally on tape) and arranged them according to some chance procedure.
- I wanted to do that sort of thing but instead taking all of the samples from a single cohesive audio file (e.g. a song), and grouping them somehow.
- This script groups the samples with unsupervised machine learning, specifically k-means clustering, where the features used are randomly selected bins from a spectrogram of the samples.
- A number of separate tracks are generated, each composed of samples all coming from the same group as determined by clustering.
- The density of samples throughout each track varies periodically based on a sine wave at a random frequency.
- Some other stuff is randomised.
- The tracks are then mixed together.
There is also a version that uses MFCCs as features, but I haven't updated that one. I don't really remember why I made it separate, or at all. I think it might not have worked as well.