“Always Listening” doesn’t have to be always listening

I saw a post recently in the Android Central forum that talked about Sensory’s technology as used by Samsung:

What makes it different from any other voice app is its part of the OS. e.g. get a call, you can say ‘Answer’ or ‘Ignore’. alarm rings, just say ‘Snooze’. You don’t have to launch an app or press buttons to do this, the phone is always active and listening. No one else does this!

It’s an astute comment but not 100% accurate. When people talk about “always listening” what they really mean is that it appears to be “always listening”. At Sensory we call it TrulyHandsfree, and the idea is that there can be certain “modes” or “windows” where it listens for specific words. Like when the alarm goes off, it listens for “snooze” etc. If you say “snooze” when the alarm isn’t going off you find it’s not really “always listening”.

Glass has a similar usage model. It’s “always listening” but for different things at different times and only for short periods of time. I put my Glass on and timed it. The OK Glass trigger window seems to last 3-4 seconds, then the next set of commands (like Get Directions to) stays on 10-11 seconds.

What’s really cool about Glass is that during those listening windows you can say other things and it doesn’t “false fire” on them. I let my wife try out my Glass, and she said “You mean I just say OK Glass and then I can say any of these things like get directions to Chef Chu’s and…woah it works!” It ignores everything it’s not listening for and picks out the things it is listening for. The technology is known as “keyword spotting” for this reason.

To save power, Hallmark’s use of Sensory’s technology kicks into gear when the product is turned on. If it doesn’t hear one of the words it’s listening for spoken within a certain time frame, it will automatically power down, and stop “always listening” until its turned back on with a button press.

Sensory recently introduced a low power sound detection technology that further cuts power consumption by having the device “always listening” in a low power mode, where it doesn’t perform speech recognition. When it hears something it quickly powers up the recognizer for further analysis. This can cut the power consumption by “always listening” but not always recognizing, down to 1mA or so.