Generative adversarial training of the noise generation model for speech synthesis using speech in noise