AI can tell what you're typing by listening to the sound of your keyboard (theregister.com)

> over Zoom, the accuracy of recorded keystrokes only dropped to 93 percent, while Skype calls were still 91.7 percent accurate.

That's interesting considering the fact that Skype messenger is a well-known of its good (if not best) audio codec.

Was this with noise filtering off? M series macs have OS level noise filtering and I believe Zoom has a built-in noise filter as well.

So a reason to add a noise gate into your effects chain. Hard to guess what you are typing from a stream of zeros.

Beware: many noise gates just attenuate the signal, that may or may not be the same as zeroing it (depending on the amount of attenuation and the level of the typing signal).

that is what Facebook and google are doing for long then?

Here is some more information about acoustic cryptanalysis, which has been used by intelligence agencies since the 1960s or so. What is new about this research is the use of AI to accomplish the side channel attack. Two decades ago the same thing made headlines when researchers used neural networks to decipher keystrokes. It is no less concerning for not bring new; I think the back story adds detail to the picture that the article only alludes to.

https://en.m.wikipedia.org/wiki/Acoustic_cryptanalysis

I guess I'm safe because I type at 100+ wpm.

What's described in the title of the article is old news. What might be novel, however, is the application of detecting typing on laptop keyboards. This type of detection was previously only reliable for mechanical keyboards.

Can this tech be used for making powerless keyboards? As in no power just a mic.

Obviously this has bigger implications than passwords, but the password discussion in the article made me realize I’ve never typed most of my passwords. I didn’t consider that as a benefit of password managers, but I’ll take it.

Good luck with me, I use 34 keys in a Colemak layout with 8 levels of layers, essentially rendering it all useless, many strokes are rolled into the next, and chorded combos will sound like one ambiguous click.