TitleConnectionist models for real-time control of synthesis and compositional algorithms
Publication TypeConference Paper
Year of Publication1992
AuthorsLee, MA, Wessel, D
Conference NameInternational Computer Music Conference
PublisherInternational Computer Music Association
Conference LocationSan Jose, CA
Abstract

Connectionist models provide rich and trainable control structures for generative algorithms. Multilayer neural networks trained by back-propagation can be used effectively to transform performance gestures into control parameters. The MAXNet neural network simulator has been enhanced by giving it the capability to construct networks from a graphic specification. This capability facilitates experimenting with networks that have architectures richer than the conventional feed forward, fully connected networks. With this new description language, controller networks have been built based on forward models for control that drive the controller by back-propagating errors through an emulator, thus reducing the search space from that of a direct inverse technique. A user can map parameters obtained from a personalized gesture space to the control parameters of a synthesis engine. Intelligent preprocessing of gestural data and perceptually based representations of sound are critical determinants of the performance of such network-based control structures. Examples include live performance control where the performer makes gestures in a low dimensional, perceptually based timbre space and controls either FM, resonant synthesis, or waveguide synthesis. (authors)

URLhttp://cnmat.berkeley.edu/publications/connectionist_models_real_time_control_synthesis_and_compositional_algorithms