What the heck are Subsonic Commands?


#1

I just found this article from Venture Beat. Interesting twist, I’d say.
google-assistant-can-be-controlled-by-inaudible-commands


#2

Speech to text engines process audio that humans can’t hear ( either above or below human hearing ). Security researchers have developed software that shifts voice commands into the inaudible range. From the voice assistant’t point of view it is just a command.

Security in the voice space is an important item that needs to be addressed, probably through voice print, but that has it’s own challenges ( biometric data can’t be changed i.e. you can’t change your voice, so if someone compromises your data it is permanent ).

Voice technology is in it’s infancy. It isn’t the first inning of the game, it is the first at-bat.


#3

Good luck trying to combat subsonic commands - they’ve been around for years (Hollywood uses them in piracy tracking, for example) and by design, they’re difficult to block… You could possibly introduce some sort of system to permit only “authorized voices”, but that then opens up its own can of technical limitations and challenges.

The only (realistic) Real World solution for the time being would be to control the environment by limiting Mycroft’s access to surrounding devices with a speaker or microphone… But if you’re going to go to that extent, one might wonder why you’re using Mycroft at all.


#4

What might be easier than “authorized voice” is frequency domain checking. Is it possible to frequency-check a “Hey Mycroft” capture and ignore it if the check says that it is below human speech range?

I may be wrong, but I’m thinking it would need a whole lot less processing power to analyze frequency than to parse out specific voices.


#5

Or maybe put some sort of high-pass filter on the mic, either in hardware, or at least in software before the audio even gets to Mycroft. Hardware would use less (well, no) extra processing power, but software would be more flexible (in cases like ones where that sort of hardware change is infeasible, or you’re grabbing audio from other (remote?) sources).

No checks needed then, because the subsonic would never even make it to Mycroft that way.