The paper discusses RaD-Net, a repairing and denoising network for speech signal improvement. The authors have improved their previous two-stage neural network model by replacing the repairing network with COM-Net from TEA-PSE. They also introduced multi-resolution discriminators and multi-band discriminators during the training phase. A three-step training strategy was employed to optimize the model. The proposed systems ranked 2nd in track 1 and 3rd in track 2 of the ICASSP 2024 Speech Signal Improvement Challenge.
Publication date: 11 Jan 2024
Project Page: https://github.com/mishliu/RaD-Net
Paper: https://arxiv.org/pdf/2401.04389
