For a little background: we're using an active listening model to listen for voice commands. As soon as Jarvis hears a command word such as "lights" within a few words of his name, he acts on it. 

Lately we've been combating false positives - during a talkative party, our system picks up a lot of things that sound like "Jarvis, lights" for instance. We're working on a number of possible fixes, including measuring the general room activity (judged by number of words transcribed in the last minute) to determine if a more complex command should be needed to command Jarvis (for instance requiring the user to say "please").