Don't typical home assistants already do voice recognition in the cloud? It's just the attention phrase ("ok Google" etc) they recognize locally

Depends how you define "biggest" but running an ML physics sim per-frame per-character in a AAA title would add up to a hell of a lot of inference.