Time-Frequency Generative Adversarial Network with Transformer for Packet Loss Concealment

With the advancement of internet technology, internet communication has become an indispensable part in our daily life. However, in actual conversations, due to the latency and jitter of the network, the data packet may be lost and have a negative impact on the quality of communication. This paper introduces a generative adversarial network with attention mechanism, which effectively restores the missing speech by the already received historical information. The generator composed of dilated convolution and Transformer module is employed to extract both local and global features from the historical information and generate the missed speech. The discriminator is composed of two multi-resolution discriminative networks for further enhancing the accuracy of the recovered speech. Experimental validation demonstrates the effectiveness of this approach in restoring lost speech, and the communication quality in the occurrence of packet loss is enhanced.