Generative Music
The final goal of this project is to create a fully end-to-end generative music model. An algorithmic musical maestro!
Version 1
I began this project as part of the HackNEHS competition, but it has expanded as an independent project unto itself. Its goal is to programmatically generate music.
What was the process?
V1: In version one of this program, which was presented in the competition, the model was a simple abstract, probabilistic model. It was similar to a Markov Chain for deciding chord progressions and pure, logical probabilities for melody generation.
V2: In version two, I worked with abstract Hidden Markov chains and n-grams. This one wasn't very successful, so I fairly quickly abandoned it.
V3: In version three, I'm working with Generative Adversarial Networks and a large midi dataset I have collected. It's currently training on an Amazon AWS p2 instance.
V2: In version two, I worked with abstract Hidden Markov chains and n-grams. This one wasn't very successful, so I fairly quickly abandoned it.
V3: In version three, I'm working with Generative Adversarial Networks and a large midi dataset I have collected. It's currently training on an Amazon AWS p2 instance.
What did I learn?
- Amazon AWS
- Had to figure out their weird p2 instances
- GAN
- Really new idea and little optimization advice available
- Keras
- Really nice frontend for theano
Results?
My team won the HackNEHS competition with this project. Otherwise, there is nothing to share with the wide world . . . yet.