Statistical parametric speech synthesis using generative adversarial networks under a multi-task learning framework