Deep Neural Network Modeling of Distortion Stomp Box Using Spectral Features

We propose a distortion stomp box modeling method using a deep neural network. A state-of-the-art method exploits a feedforward variant of the original autoregressive WaveNet. The modified WaveNet is trained so as to minimize a loss function defined by the normalized mean squared error between the high-pass filtered outputs. This method works well for stomp boxes with low distortion, but not for those with high distortion. To solve this problem, we propose a method using the same WaveNet, but a new loss function, which is defined by a weighted sum of errors in the time and frequency domains. The error in the time domain is the mean squared error without high-pass filtering. The error in the frequency domain is the generalized Kullback-Leibler (KL) divergence between spectrograms, which are given with a shorttime Fourier transform (STFT) and a Mel filter bank. Numerical experiments using a stomp box with high distortion, the Ibanez SD9, show that the proposed method is capable of reproducing high-quality sounds compared with the state-of-the-art method especially for high-frequency components.