Was performed at:
This is not a piano is a suite of pieces for sampled piano I wrote (or generated) from small snippets of text data. Learning a variable-order Markov chain from the properties of the snippets, I added variation through layering and manipulating certain aspects of the sound in an improvisational process, using my live coding environment Mégra.
The results have a slight reminiscence of, though no real connection to, some piano pieces by Eric Satie. I’ve always been drawn to the piano as an instrument, as apart from its pleasant sound it also lends itself well to reflect structural properties of music.
In the end, it’s more about the interaction and dialog with the algorithm rather than trying to beat benchmarks.
The abstract is displayed here for proof-reading and will only be part of the published proceedings, not of the final version of this web catalogue.
This is not a piano is a suite of pieces for sampled piano I wrote (or generated) from small pattern-like snippets of text data. Learning a variable-order Markov chain from the properties of the snippets, I added variation through layering and manipulating certain aspects of the sound in an improvisational process, using my self-developed live coding environment called Mégra.[^1]
The results have a slight reminiscence, though no real, material connection to, some piano pieces by Eric Satie. I’ve always been drawn to the piano as an instrument, as apart from its pleasant sound it also lends itself well to reflect structural properties of music. Thus, live coding with piano samples was a natural consequence for me (a Disklavier is typically unavailable, even though I’d love to perform the piece on one).
The piece isn’t using the state-of-the-art AI methods we have seen popping up in the last decade. The methods it uses are well known since the mid-nineties, and in fact, it’s in many ways closer to old-fashioned generative music than what is currently popular under the label “AI Art”.
Due to the small amount of data, the results are not a 1-to-1 reproduction of the input pattern, but rather an unsharp interpretation of the pattern by the language. What in other context would be perceived as a lack of data or an insufficient learning algorithm is used here intentionally, to create musical variation, which is not as much of a goal-oriented process as, say, recognizing car license plates or other typical machine learning tasks.
In the end, it’s more about the interaction and dialog with the algorithm rather than trying to beat benchmarks.1
The Mégra System - Small Data Music Composition and Live Coding Performance - https://doi.org/10.5281/zenodo.3939154 ↩