Deep Learning for Tube Amplifier Emulation

Eero-Pekka Damskägg, Lauri Juvela, Etienne Thuillier, and Vesa Välimäki

Demo page for the paper submitted to ICASSP 2019.

The following sounds are electric guitar and bass guitar sounds processed through different black-box models. The black-box models are emulating the preamplifier of the Fender Bassman vacuum-tube amplifier. 'Block' is the block-oriented model proposed by Eichas and Zölzer (2016), 'MLP' is a multilayer perceptron, 'WaveNet1' is a relatively small WaveNet model with only approx. 600 trainable parameters, and 'WaveNet2' is a larger WaveNet with approx. 30k trainable parameters. The sounds are presented with two amplifier settings: gain potentiometer at 50% and at 100%. The MLP and WaveNet models adapt according to changes in potentiometer position. For the Block model, different models have been trained for the two potentiometer settings. More details can be read from the paper.

Conference Poster