Real-Time Modeling of Audio Distortion Circuits with Deep Learning

This paper studies deep neural networks for modeling of audio distortion circuits. The selected approach is black-box modeling, which estimates model parameters based on the measured input and output signals of the device. Three common audio distortion pedals having a different circuit configuration and their own distinctive sonic character have been chosen for this study: the Ibanez Tube Screamer, the Boss DS-1, and the Electro-Harmonix Big Muff Pi. A feedforward deep neural network, which is a variant of the WaveNet architecture, is proposed for modeling these devices. The size of the receptive field of the neural network is selected based on the measured impulse-response length of the circuits. A real-time implementation of the deep neural network is presented, and it is shown that the trained models can be run in real time on a modern desktop computer. Furthermore, it is shown that three minutes of audio is a sufficient amount of data for training the models. The deep neural network studied in this work is useful for real-time virtual analog modeling of nonlinear audio circuits.