ON the study of generative adversarial network for corss-lingual voice conversion