Adversarial representation learning for private speech generation