A Time-Frequency Band-Split Neural Network For Real-Time Full-Band Packet Loss Concealment

Lingling Dai, Yuxuan Ke, Huiyong Zhang, Fengyuan Hao, Xiaoxue Luo, Xiaodong Li, Chengshi Zheng

Abstract: This paper describes a time-frequency band-split neural network for the real-time full-band speech package loss concealment (PLC) task. The proposed method is based on the generative adversarial network (GAN) structure comprised of a modified UNet generator and multi-domain discriminators. Specifically, besides a full-band branch, the generator contains an extra high-band encoder, which was designed to compensate for the high-frequency information and guide the full-band branch focusing on the recovery of the low and medium-frequency components. Our proposed method ranked the 3rd place in the ICASSP 2024 PLC Challenge, achieving a P.804 overall mean opinion score (MOS) of 3.37, a P.804 discontinuity MOS of 3.67, and a word accuracy (WAcc) ratio of 0.81.