Two-Stage Neural Network Model with Packet Loss Detection for ICASSP 2024 PLC Challenge

We proposed an audio packet loss concealment method based on a two-stage neural network model. Specifically, in the first stage, the encoder-decoder architecture is used to generate a filter in the time-frequency domain for packet loss concealment. Meanwhile, packet loss detection is conducted with the potential features from the encoder module. In the second stage, another filter is estimated with the information from the first stage to further enhance the distortion. In this model, we proposed a local dense convolution (LDC) layer to improve the ability of the conventional convolution layer without increasing computation complexity. The experiment results verified the effectiveness of the two-stage method and our proposed LDC layer. In ICASSP 2024 audio deep packet loss concealment challenge, our system ranks 5th in the evaluation of P.804 and word acceptance rate.