New open-source voice model listens nonstop and decides every 0.4 seconds whether to speak or stay silent — Blankdot