Hey everyone! This is my first Junto submission. I’ve long wanted to give it a try, and this week’s prompt spoke to me.
So here’s the setup:
When you’re using machine learning systems to generate output – whether it’s text, images, or audio – there’s often a “temperature” parameter you can set, which deeply affects that output’s character. A sample generated at high temperature will be “adventurous,” maybe even chaotic; a sample generated at low temperature will be “safe” and usually pretty blah.
When I read “a robot has the blues,” I thought immediately of an ML system operating at low temperature. A bit glum. Depressed.
Generally, you set the temperature for an entire sample, so the resulting output – a whole sentence, a big chunk of audio – is uniformly surprising or boring. But it’s also possible to adjust the temperature as the generation process unfolds, and that’s what I did here. Before I went to sleep last night, I set my computer up to generate a few dozen samples using a model trained on a dataset that leans cinematic and synth-y. Each sample was to be 120 seconds long, declining smoothly from temperature 1 to temperature 0.
This morning, I sorted through those samples, chose three that I liked, dropped them into Ableton Live, and built a short track around them. I hope you can still detect the temperature decline, modulo my own arrangement. The percussion is from a Moog DFAM.
