Exploring Algorithmic Composing from the 18th Century through Machine Learning – work in progress
Methods for algorithmic music composition have been developing since the 18th Century, a lot of composers used rules to compose music. The most famous ones are C.P.E Bach and W.A. Mozart.
The first stage of the project is focused on the algorithmic composing mode in the past.
The Kaleidacousticon is an algorithmic musical composition and manuscript printing machine. Named after a late Victorian generative music parlour game.
The dice rolls randomly selected small sections of music, which would be patched together to create a musical piece. Mozart’s manuscript, written in 1787, consisting of 176 one-bar fragments of music. It was published in 1792, by Mozart’s publisher Nikolaus Simrock in Berlin (K. 294d or K. 516f).
The physical device is inspired by typewriter and fortepiano mechanisms.
Each hammer head holds a short fragment of the music from Mozart’s.
Here are two examples generated from his score.
The algorithm uses two dice to write waltzes based on pre-composed fragments. All the fragments are numbered and can be found in a table.
After rolling 8 times, all the blanks have been filled with fragments.
Each hammer head holds a short phrase of the music. Once the music is generated, the device stamps each bar onto paper. A unique but similar tune will be presented in front of the user.
The Second Part:
The modern approach on algorithmic composing. The ground breaking exhibition: Cybernetic Serendipity mentioned really early stage of computer plotting and ink dropping method to generate music.
With the development of machine learning. The logic and rules become more abstract and complex.
Rosemary Brown was an English composer, pianist and spirit medium who claimed that dead composers dictated new musical works to her. She wrote down all the new tunes composed by the dead composers in their right period. Does our AI explore the same thing?
My approach on machine learning composing is based on MusicVAE a model from Magenta built upon tensorflow. One of the feature of the MusicVAE is it could generate the linking part between the starter and end points. Here is one example of my process on generating variations for the most well-known Scottish Baroque music collected by Niel Gow
Set up the first measure as the starter point, the second measure as the ending point. Develop the parts between with machine learning model.
As the structure of the tune and the chord progress, the tune need to be manually rearranged into its mode: A1B1A2B2A3…..
Here are some of the variation set generated by the model