Towards Robust And Practical Neural Video-Conferencing