Probably only the voice is local. Natural voice requires a voice synthesis model. The “reasoning” is probably offloaded to a cloud service.