Predicting Transformed Audio Descriptors: A System Design and Evaluation

TitlePredicting Transformed Audio Descriptors: A System Design and Evaluation
Publication TypeConference Paper
Year of Publication2010
Conference NameWorkshop on Machine Learning and Music (ACM Multimedia 2010)
AuthorsColeman, G., & Villavicencio F.
Conference Start Date25/10/2010
Conference LocationFirenze, Italy
Abstract
We propose and present an example system design for predicting changes in perceptually relevant audio properties under the eff ects of common musical and sonic transformations. By building these predictive models, we may facilitate descriptor-driven control of eff ects while avoiding queries to the transformation itself. In this study we model spectral descriptors of a limited class of sounds under the resampling transformation with Support Vector Regression (SVR) and report on the accuracy of the predictions, with an emphasis on performance as a function of model parameters. On a test set of resampled inputs, the statistical model predicts an output lter bank within 3-4 times the mean absolute error of a comparable analytical model.
preprint/postprint documentfiles/publications/mml10.pdf
intranet