Updated: about 5x faster recognition, although accuracy is still the same. The problem was the "loading" animation: as I couldn't figure out how to do a partial redraw, it was redrawing the entire screen, 20 times a second while recognizing, leaving little CPU time free for the actual recognition. I still haven't figured out partial redraws, so for the time being Saera shows no visual indicator while running. But you should generally get a response within a few seconds.

Also added: times. Ask Saera "What time is it?" or "What is the time in <city>" where <city> is a reasonably large city like london, paris, tokyo etc.