The Alexa hardware is cheap and doesn't have a lot of computational horsepower, so you can get better accuracy by doing the speech recognition on the server side.