In recent years, automatic music-driven choreography has become a highly challenging problem to be solved. In this paper, we propose a music-driven choreography system based on conditional generative adversarial networks. First, a dataset MF-DS integrating MFCC features and Dancing Skeletons extracted from Japanese dancing videos is built by ourselves for the study. The MFCC features are extracted based on music beats, and the dancing skeletons are detected based on the image frames of a video. In the training, we use a generative adversarial network to train the music-driven choreography system. The generator integrates residual blocks into fractionally stridden convolution, and the discriminator involves conventional CNNs. Two indicators called beat loss values and choreography diversity values are proposed to evaluate three learning models in the experiments. Finally, we validate that the three models with the best epochs have the near-zero loss for the generator and discriminator, thereby generating stable skeletons and presenting choreography diversity.