Learning Parameter Spaces in Neural Modeling of Audio Circuits

Abstract

This thesis studies black-box virtual analog modeling formulated as a machine learning sequence modeling task within the category of supervised learning problems. The focus is on learning scenarios where the modeling targets have multiple user controls, and the aim of the thesis is to evaluate how the properties of the training datasets affect the generalization of the learning algorithm.

To study the problem, three nonlinear analogue sound processors were modeled using a recurrent neural network consisting of a Gated Recurrent Unit and a fully-connected output layer. For each target device, two groups of datasets, seven in total, were constructed, using SPICE simulations of the targets. The difference between the datasets is in the density of the sampling grid used for setting the user controls of the targets, as well as in the number of input/output pairs corresponding to each distinct value of each of the controls.

For the targets considered during the study, the sparsest sampling grid using only three possible values for each of the user controls was found inadequate for the models to generalize over the testsets used for evaluation. Increasing the sampling density was seen improving the model performance in most cases, with some targets also portraying clear advantages with increasing the number of input/output pairs corresponding to each distinct value of the user controls. According to the study, a sampling grid with five points would appear as a good baseline for training neural networks on targets with multiple user controls when no further investigations in the sampling density can be afforded.

For future work, the experiments could be extended to include global scaling of the dataset size while keeping the constraints for sampling the parameter spaces, as well as combining the data generation and training procedures to a single loop, allowing for potentially infinite variety within the datasets.