Not sure what you mean exactly but this seems wrong. These devices do hotword (and speaker) recognition largely locally, but speech to text is done server side.