The best way would be to have minimal hardware detectors as possible that can detect the subtle bone vibrations that we make when subvocalizing, the heavy lifting would be done by software AI is probably necessary for recognition of individual's distinct subspeech.