A Vacuum-Tube Guitar Amplifier Model Using Long/Short-Term Memory Networks

In a previous paper the authors use a recurrent network in the form of a NARX architecture to model the nonlinear behavior of a vacuum-tube guitar amplifier and its effects as applied to an electric guitar signal. The use of recurrent networks is important in the vacuum-tube modeling effort as the nonlinear dynamic responses are thought to be an important component in achieving the distortion that creates the sound treasured by rock guitarists and the fans of rock music. In this paper, we report on our current experiments evaluating the use of long/short-term memory (LSTM) units for modeling the dynamic nonlinear characteristics of a vacuum-tube guitar amplifier. It is our view that since LSTM units incorporate information from both recent and distant changes in state, they are useful for modeling the amplifier characteristics that produce the desired musical timbre. Our experiments involve investigating differing combinations of the hyperparameters of a deep learning network including the number of hidden layers, the number of units within each layer, and the type of unit within each layer. Model performance is evaluated using mean-square error and aural comparisons. Once the best architecture is determined, efforts will be put forth to minimize the hardware requirements for the implementation of the feedforward network.