Cross-Utterance Conditioned VAE for Speech Generation