Musical Instrument Mapping Design with Echo State Networks Chris Kiefer Goldsmiths College, University of London
Long Paper
Echo State Networks (ESNs), a form of recurrent neural network developed in the field of Reservoir Computing, show significant potential for use as a tool in the design of mappings for digital musical instruments. They have, however, seldom been used in this area, so this paper explores their possible uses. This project contributes a new open source library, which was developed to allow ESNs to run in the Pure Data dataflow environment. Several use cases were explored, focusing on addressing current issues in mapping research. ESNs were found to work successfully in scenarios of pattern classification, multiparametric control, explorative mapping and the design of nonlinearities and uncontrol. \emph{Un-trained} behaviours are proposed, as augmentations to the conventional reservoir system that allow the player to introduce potentially interesting non-linearities and uncontrol into the reservoir. Interactive evolution style controls are proposed as strategies to help design these behaviours, which are otherwise dependent on arbitrary parameters. A study on sound classification shows that ESNs can reliably differentiate between two drum sounds, and also generalise to other similar input. Following evaluation of the use cases, heuristics are proposed to aid the use of ESNs in computer music scenarios.
