Using conditional restricted Boltzmann machines to generate timbral music composition systems

music composition;conditional restricted Boltzmann machines;human-computer interaction

Junokas, Michael J.

关键词: music composition;    conditional restricted Boltzmann machines;    human-computer interaction;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/101456/JUNOKAS-DISSERTATION-2018.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Machine-learning models have been successfully applied to musical composition in a variety of forms, including audio classification, recognition, and synthesis. The capability of algorithms to learn complex musical elements allows composers to more deeply investigate the development of their aesthetic. Coupled with the history of interdisciplinary solutions found in computer music and system aesthetics, this capability has led to an exploration of the integration of machine learning and music composition. Composition systems that take advantage of this integration have the opportunity to be connected with algorithms in theory, application, and art.In my systems, conditional restricted Boltzmann machines (CRBM) synthesize musical timbre by learning autoregressive connections between the current output, an abstracted non-linear hidden feature layer, and past out- puts. This provides a creative space where composers can synthesize audio spectra in collaboration with machines, defining novel creative systems that explore compositional material in an abstract, non-linear paradigm.By implementing CRBMs in timbral-synthesis composition systems, I provide concrete support that such an integration advances art through the exploration of machine learning. I demonstrate this in a variety of audio synthesis experiments validating the capabilities of two algorithmic structures to synthesize and control timbre: a single layer conditional restricted Boltzmann machine (CRBM) and a single layer factored conditional restricted Boltzmann machine (FCRBM). I start by accurately synthesizing specific instrumental timbres and different musical pitches, demonstrating the aural capabilities of directly using the algorithms. I then build from these experiments, creating a set of compositional utilities that provide the composer with a rich pallet to provoke aesthetic introspection. These compositional utilities are then implemented in two music composition systems that synthesize and control timbre in application, where the algorithms themselves are designed and manipulated as a means to realize artwork.Through the creation of music composition systems that are able to accurately synthesize and control musical timbre, I demonstrate these models have the capability of provoking the aesthetic introspection of composers. The resulting systems show the power and potential of integrating music composition and machine learning, endorsing an interdisciplinary approach to the development of art and technology.