Deepjazz

Deepjazz is an open source project that generates jazz music with deep learning. Built by Ji Sung Kim of the Princeton University Department of Computer Science in 36 hours at a hackathon, it trains a two layer LSTM network on a supplied MIDI file and then produces new jazz passages in the style of that input. The code is written with the Keras and Theano deep learning libraries and is published on GitHub, with downloadable zip and tarball archives on the project page. Its audience is developers, students and researchers interested in algorithmic music composition and sequence modelling rather than musicians looking for a finished composing tool. The project page notes that it is used as reference material for the course Interactive Intelligent Devices at the University of Perugia, and that it builds on preprocessing code from Evan Chow's jazzml project, used with permission. Output quality depends on the MIDI file used for training, and the workflow is run from code rather than through an interface. Deepjazz is free to download and use. There is no paid version, hosted service or published pricing. The project should be treated as a compact reference implementation rather than maintained software. It was produced as a hackathon build, it depends on Theano and an early Keras release, and running it today generally means recreating an older Python environment or porting the model to a current framework.