A Fullband Neural Network for Audio Packet Loss Concealment

Audio packet loss, a result of missing or irregular packets, significantly degrades speech quality in voice communications. The ICASSP 2024 Audio Deep Packet Loss Concealment Challenge seeks to drive research to address this issue. In this paper, we introduce a fullband neural network for PLC. Our system is based on that of the previous challenge’s winner and extends it for training and inference at 48kHz. We improve P.804 Discontinuity and Overall scores by 1.02 and 0.84 respectively, while only slightly decreasing word accuracy, placing fourth in the challenge.