Not very good, unfortunately. There is pocketsphinx which sort of works but not nearly as well as the online ones, and some other research projects that are very hard to even set up.

A while ago there was an article on here about a free (but not open) recognition engine that worked on a raspberry pi, but I forgot the name - the founders were hanging around here back then, maybe they can chime in? I haven't had a chance to try it though.

Kaldi can produce excellent results, but is a pain to build, configure, and train.

Couldn't building, configuration, and training be done separately and the results distributed? Or is the data for training publicly available?

Sorry, but I feel that the community can do a big catch up in this area.

There was (or is) one called Mycroft which had promised big things but didn't deliver a compelling recognition and synthesis, although the recognition seems better than synthesis from what I have seen. They did open source much of the work at least initially. There is a video showing an example of how bad the UX was for a simple question asking about beans https://youtu.be/D5J7vVQNkCw.