5 ms·

Sending voice data constantly is easy to catch on the user's side, by simply monitoring the traffic. So once they got caught, it is going to cause them a huge f

Sending voice data constantly is easy to catch on the user's side, by simply monitoring the traffic. So once they got caught, it is going to cause them a huge fortune and completely meltdown in terms of customer trust. And on the technical side, this makes very little sense either, they only needs to have the Alexa keyword detected then went through the backend ASR service, otherwise the cost is going to be overwhelming.

uptown 8y ago
Why would you need voice? Just translate to text and send that. It’d be a very very small payload.
- teraflop 8y ago
  The Alexa hardware is cheap and doesn't have a lot of computational horsepower, so you can get better accuracy by doing the speech recognition on the server side.
- oh-kumudo 8y ago
  The Alexa device itself is rather cheap and the running a relative huge model on 'edge' device is still fairly expensive and remains an open question. Again it falls into economic side of things, it makes sense to adapt a layered model in this case, to delay more expensive processing to the server side.
- kabdib 8y ago
  A defense lawyer would have a ton of fun poking holes in the accuracy of speech-to-text, especially on a small device. (You could easily determine if the device is doing this kind of processing by looking at power draw or thermals. This stuff takes a lot of crunch). Amazon isn't stupid. They know that if a back-door is found in these devices, or if they have bugs that allow them to be compromised, that they are toast in the market. I'd be surprised if they didn't have internal briefs and a bunch of planning on many scenarios like this. Amazon has almost certainly made the most secure device it can, knowing it will be under intense scrutiny, and will have prepped its legal teams appropriately.
  - ams6110 8y ago
    My phone, a several-year-old Moto G, a middling phone even when new, does a decent job at text to speech without noticable thermal or power issues. And I don't have mobile data so it can't be that it's offloading it to a server.
Cyph0n 8y ago
Well, the recording could be near-continuous, but compressed and stored until a user makes a request. Therefore, you’d need to record the amount of data transmitted over a certain time period and set an upper bound based on that. Of course, if you could install a custom cert on the device, you could just MiTM the connection and see exactly what is sent...
- lima 8y ago
  Technically possible? Sure. However, it just makes zero economical sense for them to do so. Sooner or later, they would get caught (reverse engineering, ex-employee, ...). Now, if the device were to get compromised, it would be an entirely different story (and this is why I don't have one, despite trusting Amazon in that regard).

uptown 8y ago

Why would you need voice? Just translate to text and send that. It’d be a very very small payload.

teraflop 8y ago
The Alexa hardware is cheap and doesn't have a lot of computational horsepower, so you can get better accuracy by doing the speech recognition on the server side.
oh-kumudo 8y ago
The Alexa device itself is rather cheap and the running a relative huge model on 'edge' device is still fairly expensive and remains an open question. Again it falls into economic side of things, it makes sense to adapt a layered model in this case, to delay more expensive processing to the server side.
kabdib 8y ago
A defense lawyer would have a ton of fun poking holes in the accuracy of speech-to-text, especially on a small device. (You could easily determine if the device is doing this kind of processing by looking at power draw or thermals. This stuff takes a lot of crunch). Amazon isn't stupid. They know that if a back-door is found in these devices, or if they have bugs that allow them to be compromised, that they are toast in the market. I'd be surprised if they didn't have internal briefs and a bunch of planning on many scenarios like this. Amazon has almost certainly made the most secure device it can, knowing it will be under intense scrutiny, and will have prepped its legal teams appropriately.
- ams6110 8y ago
  My phone, a several-year-old Moto G, a middling phone even when new, does a decent job at text to speech without noticable thermal or power issues. And I don't have mobile data so it can't be that it's offloading it to a server.

Cyph0n 8y ago

Well, the recording could be near-continuous, but compressed and stored until a user makes a request. Therefore, you’d need to record the amount of data transmitted over a certain time period and set an upper bound based on that. Of course, if you could install a custom cert on the device, you could just MiTM the connection and see exactly what is sent...

lima 8y ago
Technically possible? Sure. However, it just makes zero economical sense for them to do so. Sooner or later, they would get caught (reverse engineering, ex-employee, ...). Now, if the device were to get compromised, it would be an entirely different story (and this is why I don't have one, despite trusting Amazon in that regard).