Edited might be the wrong word. Real time guesses each utterance as it is spoken. Final version is looking at the last few seconds of utterances which give it context. The extra context means it can guess better and will fix some words. Last few seconds is actually just the sound between pauses in speech.