Plenary: A review of on-device fully neural end-to-end speech recognition and synthesis algorithms