I wonder, if using camera to read cues from lips and check if user watches on screen (as in addressing device) as a secondary channel to mic could improve situation.