This seems like a step backwards from just seeing a real persons face on a regular screen :(

The demo for that looked pretty rough. The user's facial expressions looked unnatural, and the lip movements were poorly synced to the audio. I guess there's only so much an "encoder/decoder" model can make up with limited input data.

That explains why every other user was a floating head; that would let them generate the same.

wondering how they deal with the issue that the face starts moving before we start making sound. I guess the network latency will help.