privacy

There is a "convenient" presumption in the article, and in what I'm replying-to:

That audio is uploaded to big-tech, after someone speaks near their phone..

Why TF would they need to do that, if they can cram a "hey google" audio-to-logic routine into mere kilobytes, within their neuro-DSP??

The Active Listening system mentioned in the article needs to:

1. have the onboard ai listen for voices,

2. differentiate significant items ( for whatever values of "significance" they want to manipulate )

3. upload .. what, a YAML file, .. of spoken keywords, & maybe an identification of if it was the owner or not who spoke them?

Something like that..

4. NOT react to vocal bots, or other idiocy ( a recording repeating something endlessly ought be immediately-ignored, as .. from the Active Listening manipulation-system's perspective .. mere jamming )

The "we aren't seeing audio being uploaded" is a fucking red-herring, & any competent geek nowadays, who understands that the neuro-DSP chips in our phones can accomplish stupidly capable term-recognition on near-zero code, when in active-sleep ( or whatever that energy-state is called, screen off, but things still running in the background ), and that's all they need.

You know how in Google AdWords you pay for specific keywords?

Guess what: the same system could easily be implimented within that subsystem which says "Sorry, the microphone is turned off", when you say "Hey Google" after turning off the mic..

it's ALWAYS listening, it uses tiny power, & when it matches, then it activates other subsystems..

Selling that ads can be show on phones/tablets that someone recently spoke a keyword for, would be idiot-simple for Google or Apple to impliment, & profitable-as-hell to boot.

IF the "Hey Google" audio prompt can work in a specific energy-configuration of a device, THEN this could too, as it's the same subsystem.

https://github.com/KoljaB/RealtimeSTT

as an example of something related to what I'm talking about, but not identical..