GAN-Based Time-Domain Packet Loss Concealment Method with Consistent Mapping Approach

In the speech communication over the internet, due to the latency and jitter of the network, the data packets may be lost and result in a degraded user experience. Therefore, the packet loss concealment (PLC) technology plays an important role in recovering the lost information. In this paper, a PLC method only using few historical information is proposed based on generative adversarial network (GAN). In this method, the generator is composed of the improved gated attention unit (IGAU) and temporal convolution network, which can fully exploit the available time-domain information and recover the lost speech signal. Specially, a consistent mapping approach is devised to enhance the quality of the generated speech. In addition, two additional loss functions are used to assist in the training of network. The experimental results show that the proposed method has better performance in terms of both perceptual quality and speech intelligibility at different packet loss rates.