A Time-domain Packet Loss Concealment Method by Designing Transformer-based Convolutional Recurrent Network

Due to the network constrain, the packet loss is inevitable in the real-time speech communication. The packet loss often leads to the short interruption of the voice communication, which seriously affects the quality of speech communication. In this paper, a transformer-based convolutional recurrent network (CRN-Trans) is designed for packet loss concealment (PLC) in time domain. In the CRN-Tans, the convolutional layers are used to extract the high-level feature of speech signal in each frame. The transformer and long short-term memory (LSTM) are combined to model the sequential information along the time dimension. Furthermore, using the self-attention mechanism, the transformer is used to compute the similarities between different positions in the long sequence and dynamically generate the weights for different positions. Thus, the lost packets can be better predicted by explicitly providing attention weights for the previous frames. The experimental results show that the proposed method has better performance in perceptual evaluation of speech quality (PESQ) and short-term objective intelligibility (STOI).