That is, superhuman ability to turn transcripts into text and a man behind the curtain makes a speech UI as good as the man behind the curtain.<p>The man behind the curtain can stop you when you get it wrong and ask questions, thus the man is a usable "front-end" if it gets words right 92% of the time. A transcription bot might get 95% of the words right, but a 1 in 20 word error rate would mean a 50% or so chance to handle a 10 word sentence.