As we just learned with Yahoo, though, once the ML models have been trained, they can be disseminated and used without the need for "cloud-scale" data or compute resources.

And, for better or worse, Dragon's text-to-speech is pretty damned good after a rather minimal amount of training.

I don't think there's anything stopping voice and intent recognition from coming back to our personal machines other than the ability to keep making money from having it come up to the cloud.