Abstract
Listening tests are often an essential part of sound design but can be resource intensive to carry out. Where there are many degrees of freedom in the sound design parameters, the ‘curse of dimensionality’, means that the number of trials required to reliably understand the impact of a particular design variable increases exponentially with increasing number of parameters. When there is a particular design goal in mind (e.g. maximum audibility, pleasantness, etc.) this can be somewhat mitigated by using efficient optimisation techniques with online sound generation during listening tests -- whereby a black box optimiser iteratively moves the parameters towards those which produce the desired percept. We show in a pilot study that this approach can be improved yet further by first using dimensionality reduction for the synthesis parameters prior to performing the listening test. This allows sound designers to use fewer testing resources when optimising for a particular percept.
Keywords
sound design, crowdsourcing, dimensionality reduction, human computation
DOI
https://doi.org/10.21606/drs.2022.729
Citation
Barker, T., Vieira, J., Pereira, F., Marques, R., and Campos, G. (2022) Listening tests for sound design: Faster optimization through lower-dimensional parameter spaces, in Lockton, D., Lenzi, S., Hekkert, P., Oak, A., Sádaba, J., Lloyd, P. (eds.), DRS2022: Bilbao, 25 June - 3 July, Bilbao, Spain. https://doi.org/10.21606/drs.2022.729
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License
Conference Track
Research Paper
Included in
Listening tests for sound design: Faster optimization through lower-dimensional parameter spaces
Listening tests are often an essential part of sound design but can be resource intensive to carry out. Where there are many degrees of freedom in the sound design parameters, the ‘curse of dimensionality’, means that the number of trials required to reliably understand the impact of a particular design variable increases exponentially with increasing number of parameters. When there is a particular design goal in mind (e.g. maximum audibility, pleasantness, etc.) this can be somewhat mitigated by using efficient optimisation techniques with online sound generation during listening tests -- whereby a black box optimiser iteratively moves the parameters towards those which produce the desired percept. We show in a pilot study that this approach can be improved yet further by first using dimensionality reduction for the synthesis parameters prior to performing the listening test. This allows sound designers to use fewer testing resources when optimising for a particular percept.