Audio super-resolution with deep neural networks

Abstract

This thesis reports various attempts at applying generative deep neural networks to audio for the task of recovering a high quality audio signal when given a low sample rate signal. Our experiments show that deep networks are able to discover patterns in speech and music signals by working in both time and frequency domains jointly. Such a network structure outperforms other methods that work either in the time domain or frequency domain exclusively. In our evaluations with speech signals, our method outperforms a time-domain only method by Kuleshov et. al. by 1.4 dB for 4x and by up to 2.0 dB for 8x upsampling.