Hacker News Reader

Comment by gumby | Hacker News Reader

This is wonderful news.

My gf has one in her living room. It is only used to play music, set a timer (usually watch is used) and adding things to a (third party) grocery list. Plus it's listening all the time to what goes on.

I had one alexa device: a scanner that would read a barcode and add it to a shopping list (or you could talk to it -- ugh, had to keep it in a drawer). It was handy as I ran out of milk to just scan it and forget about it, so that I would see "milk" when at the shop. But they discontinued that device.

View on HN

I'm curious, did you actually verify that "it's listening all the time to what goes on"? The Echo devices are supposed to only turn on full power and start listening when they detect the activation word ("Alexa" by default).

How do they ‘detect’ the activation word without listening? Seems to me that is tautologically impossible. The definition of detect and listen are identical here. How do you detect sound waves without listening to them?

Listening vs listening, transcribing, and sending home

Amazon has a white paper on how it's supposed to work. According to the paper, yes, it's technically listening to audio, but storing it only in a temporary RAM buffer until it detects the wake word. I'm personally mostly interested in what gets recorded or sent to the cloud.

"Echo devices use on-device keyword spotting designed to detect when a customer says the wake word. This technology inspects acoustic patterns in the room to detect when the wake word has been spoken using a short, on-device buffer that is continuously overwritten. This on-device buffer exists in temporary memory (RAM); audio is not recorded to any on-device storage. The device does not stream audio to the cloud until the wake word is detected or the action button on the device is pressed. If it does not detect the wake word or if the action button is not pressed, no audio is sent to the cloud."

https://d1.awsstatic.com/product-marketing/A4B/White%20Paper...

They could easily push custom firmware to "persons of interest" (ever criticize your government?) and nobody would ever know.

Yes. But that's also true of laptops, tablets, cell phones and lots of other items in our houses.

And that's why computers used in classified areas have their cameras and microphones physically disabled.

This has been done with cars with the inbuilt microphone for onstar.

I'd imagine a similarly way to how car cameras work; they buffer for a few seconds, and if they detect a collision, they then save that buffer instead of discarding it.

Usually a triage system. A low-power voice detection might run on the unit itself, capable of detecting speech but not recognizing it. Another NN might be running on the device itself that can only categorize the activation words. Finally they wake the network interface and send all the data off to Amazon for a datacentre to run the inference.

They tell you that they will sometimes listen when they don't hear the wake word (or hear a near miss) and report back home what they heard. It's closer to "listening all the time" than I am comfortable with.

It has to listen to detect the activation word, doesn't it? -- therefore, it's listening all the time.

The speaker also starts recording a few seconds before a command is issued, meaning there’s likely more private information in the recordings than customers are aware of. from https://www.vox.com/the-goods/2018/11/12/18089090/amazon-ech...

https://www.nytimes.com/2016/12/28/business/amazon-echo-murd...

Those two links are from two different murder cases for which Alexa/Amazon data was subpoenaed. There are probably others.

Google will let you download your entire voice activation history. That's always an interesting listen.

they really ought to delete the data after its processed

I think that semantic discussion on what constitutes as "listening" is not as interesting as what it actually does with the signal received from the microphone. I posted a description of that in this thread earlier.

Have you actually verified that it isn't?

Which assumption is the safer one?

Safe assumptions are great for managing personal privacy, but in my view, one should be careful not to present them as facts in discussions. E.g. I use privacy shields on cameras to mitigate risk of eavesdropping, but I don't claim it's actually happening. People have also reverse engineered Echo devices and analyzed their network traffic, and when someone claims the devices are "listening to everything", I'm interested in whether the claim is really based on such analysis.

Reply on news.ycombinator.com