Fig. 5
From: Improved Wasserstein conditional generative adversarial network speech enhancement
The histogram of update value of the loss function for two models. a The histogram of the update value of the SEGAN generator. b The histogram of the update value of the improved WCGAN generator. The values in b mainly range from 0 to 2, and the uneven distribution is easy to cause insufficient learning in network. Relatively, the distribution of the value in a is uniform which ranges from − 4 to 4, which indicates that the network learning is full and the system performance is better