Enhancement of Coded Speech Using Neural Network-Based Side Information

Abstract

Audio codecs generate notable artifacts when operating at low bitrates, which degrade the quality of the coded audio significantly. There have been several approaches to enhance the quality of decoded signals with and without side information. While pre- or post-processing approaches without side information can be applied directly to existing systems without modifying codecs, approaches utilizing side information can further enhance the performance while maintaining backward-compatibility with existing codecs. In this paper, we propose a method to improve decoded signals using neural network-based side information. A neural network in the transmitter side that generates the side information and another neural network in the receiver side that estimates the log power spectra (LPS) of the original signal from the decoded signal and the side information are jointly trained to accurately reconstruct the original signal. In the same line with the analysis-by-synthesis, the neural network that generates the side information in the transmitter side takes not only the LPS of the original signal but also the LPS of the decoded signal as the input by decoding the encoded bitstream at the transmitter side. Experimental results show that the proposed audio codec enhancement scheme using neural network-based side information outperformed the audio codec enhancement without side information for the same codec operating at higher bitrates.