"The only downside of the Voice Recognition API is that it requires that you be online, probably because it does the actual recognition on a remote server. I’m not sure why since desktop computers are definitely powerful enough to do it all locally,"

Likely 2 reasons:

1. Local software would be easier to reverse engineer.

2. Local software would be more difficult to update with new algorithms.

Possible issues of tuning for one language vs multiple languages in local software might be a bear too.