Music Upscaling Using Convolutional Neural Networks

Abstract

Audio upscaling with generative neural networks has been studied in the fields of super-resolution and speech bandwidth expansion. Previous approaches have worked well for speech, but not for music. We propose a convolutional neural network approach with a novel dilated and residual architecture for this domain and an additional refinement method which outperforms the cubic spline baseline when upscaling music according to a spectral distance error metric.