Speech Codec Enhancement With Generative Adversarial Networks

Abstract

In order to solve the problem of voice quality degradation after the encoding and decoding of various encoders, we propose a processing method for the back-end of various encoders based on a generative countermeasure network. This method uses a generative adversarial network to learn the relational mapping of the speech time domain before and after encoding by the encoder, and conducts training through end-to-end training to restore the quality of the speech encoded by the narrowband encoder. We selected the encoders G.726, G.729 and G723.1 to form a training set with the original speech respectively for training, used a test set composed of speakers that did not appear in the training set for model evaluation to verify the feasibility of the model. It can be seen from the experimental results that the generative adversarial network model we explored improves the encoded speech quality.