In normal human interaction you can see people's reactions to what you're saying and doing within microseconds of taking an action, you can subtly and automatically adjust your presentation, your pace and your expressions to build greater rapport. Over video calls you need to introduce artificial pauses to make sure that everyone has caught up, and it's really jarring compared to being face to face.
Since all the character assets are downloaded locally, the only visual data that needs to be sent across network is the positioning of facial features and hand movements - which is a much smaller payload than streaming video. More akin to the latency you'd see playing Call of Duty or something.
I could see this experience being higher fidelity than video calls, especially in poor network conditions.
The application is taking the positional data of your headset/controllers and translating that into how your avatar moves around. There isn't any video analysis going on.
However, you don't notice it in other people's movements until there is a mismatch between social clues, Even then you have up to 250 ms (sometimes more, depending on the pace of the conversation)
In practice that means you can do an atlantic call without really noticing it.
the problem with VC calls is that because video needs to be matched to audio, the audio is buffered to make sure its lined up, this means that you have a 250ms+ delay plus any time to switch presenters.