LONG SHORT-TERM MEMORY

Google Brain’s Magenta — a research project exploring the role of machine learning in the creation of art and music — formed the basis of the compositional process. A dataset comprising a large body of MIDI data drawn from existing Metaludios was used to train several Magenta models. Training lasted approximately 50 hours, after which hundreds of MIDI files were generated and three were selected.

To improve readability, these files were re-notated in Sibelius (notation software), with phrasing, pedalling (based on MIDI sustain data), and some expression marks added. The music was also redistributed between the hands in a more pianistic, practical way. Apart from this, no modifications were made to the original machine-generated data in terms of rhythm, tempo, pitch, dynamics, or pedal information.

As a composer, I am both fascinated and mystified by the result. I clearly hear reminiscences of other Metaludios, although nothing is ever repeated literally. The algorithm has evidently learned from the dataset and can generate material “in my own style.” I believe that machine learning has great potential as a creative tool for composers.