Predicting transformed audio descriptors: a system design and evaluation

• Authors: Graham Coleman;Fernando Villavicencio
• Venue: Proceedings of 3rd international workshop on Machine learning and music
• Year: 2010

We propose and present an example system design for predicting changes in perceptually relevant audio properties under the effects of common musical and sonic transformations. By building these predictive models, we may facilitate descriptor-driven control of effects 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 filter bank within 3-4 times the mean absolute error of a comparable analytical model.