HIDDEN STATES

Hidden states uses Google Brain’s Magenta, a research project exploring the role of machine learning in creation of art and music, as the basis of the compositional process for this metaludio. A dataset consisting of a large body of MIDI data drawn from existing metaludios was used to train several Magenta models. The training lasted for approximately fifty hours. Hundreds of MIDI files were generated, of which three were finally selected. To improve readability, these MIDI files were re-transcribed into Sibelius notation software, adding phrasing, pedalling (based on the MIDI sustain information), and some expressive marks. The music was redistributed between the hands in a more pianistic, practical way, but otherwise no modifications were made to the original machine-generated MIDI files. As a composer, I am both fascinated and mystified by the result. I can clearly hear reminiscences of other metaludios, although nothing is ever repeated literally. It is evident that the algorithm has learnt from the dataset and it is able to generate material in my own style. I believe that machine learning has a great potential as a creative tool for the composer. Book V features the last of the three chosen MIDIs. The other two pieces (titled Long Short-Term Memory and Forget gate), together with Hidden States make up a three-movement work entitled AI Sonata. Long Short-Term Memory and Forget gate will be part of Book VI of metaludios.