DNN-based Decoded Audio Enhancement using Vector-Quantized Side Information

Abstract

Audio codecs have been studied for several decades to achieve better perceptual quality with lower bitrate. The perceptual audio codecs may induce audible coding artifacts when encoding audio at low bitrates. To solve these problems, codec post-processing methods have been studied. Recently, deep neural networks (DNN) based approaches have been studied to enhance the quality of the decoded signals. If the number of patterns of residual, which is the difference between the original signal and the decoded signal is limited, the decoded signal can be improved by learning the patterns as a codebook. In this paper, we propose the neural network-based decoded audio enhancement using side information. Our experiments on TIMIT and VCTK dataset demonstrate that the proposed method is more efficient than DNN-based enhancement without side information at a higher bitrate.