Multi-Stage Training For Cross-Domain Full-Band Audio Packet Loss Concealment

This paper introduces a multi-stage full-band packet loss concealment (PLC) method that incorporates both time-domain and frequency-domain learning. Specifically, the time-domain model is elaborately designed to generate the missing packets, while the frequency-domain model aims to remove discontinuity and noise. The proposed system achieved the 1
^st
place ranking in the ICASSP 2024 PLC Challenge, demonstrating superior performance with a mere 20ms algorithm latency.