With enough training data I’m sure it can be done... I’ve seen CNNs that fill in 3D scenes and animate them from two images. I would guess this was a simpler problem?
I -guess- CNNs can look at e.g. reduced frequency range recordings (like phone calls), and attempt to reconstruct them. However this seems like an arduous mountain to climb, as people's voices are unique. So are their environments and signal chains. I really doubt that something that generalized would work very well at reconstructing a specific person's voice and recording.
This also gets into the problem that it would be constructing a new reality, not recreating it.
I have thought about the constructing a new reality thing - I wouldn't be surprised if models ended up being trained to misspeak words which could get confusing...