It is not just a sell. I build Neural Networks as part of some work I am doing, and the kinds of multi-layer deep networks that do the recognition just don't perform well on small devices. Also the best systems often have many different trained networks and specialized routing to find the best network for your accent, cadence, noise level, etc. Having all of these weights on your local device, running all the tests, running through any of the required preprocessing, and asking the network what the answer is, converting the results, etc. is going to be too much for your phone or low powered device. While there might be other reasons for them to send data to the cloud, to get the accuracy you really need beefy servers (potentially with GPU's or something more exotic like FPGA's), and potentially multiple servers working in concert.