Thursday, April 12, 2018

Voice data bandwidth, a quick test, could Alexa listen all the time even if they wanted too?

Seeing a lot of posts about Amazon seeking a patent for voice sniffing which seems to imply they either do have the capacity on the device to recognize a whole list of trigger word or they plan on streaming all your voice to the cloud to be processed for key words. For example "Amazon patent reveals 'voice sniffer algorithm' that could analyze conversations". My response was:

These are the same devices that supposedly do not have the ability to let us select our own trigger word but will be listening for variations of I like or hate that? Does not add up. In theory they could stream all your voice to the cloud servers but that quickly starts taking up some serious bandwidth and server resources. 

Plus for all that, all they are going to find is I love my pets and hate slow computers and the cable company. No news there. Given the false triggers both my Alexas and Homes get they would build more of a profile of the characters I watch on TV than me. They would have to sort the whole speaker recog first. Alexa still is having big issues with that. Being able to read your emotional state is simple be comparison and they are still working on that. Then there is the whole I'm commenting on something I'm looking at instead of what Alexa last heard too.
Most likely here they are looking to cash in if someone attempts such a silly thing.

But as part of that I tried to find just how much data a request takes. I could not find an actual spec online so I asked my Echo Show and one of my Google Minis the time. The Mini used 254.5 KB processing the request and 26.7 KB while I was talking to Alexa. Alexa used 213.6 KB processing the request and 0 KB while I was talking to the Mini. By comparison my Roamio TiVo used 12.9 during both exchanges. That was way more than I expected and puts this even more in doubt. It takes roughly a second to say "what time is it" so those are pretty close to the kbps rates the devices sending your requests at. So say 5 of them streaming continuously would swamp the 1 Mbps upload speed of many US ISPs. Granted a lot of ISPs are advertising "up to" 5 Mbps upload now but that is still a lot especially if you are using the ISPs router provided WiFi. Then there is the other end where a server now needs to handle not only usual requests, say a 30 a day. That server has gone from processing and average of 74 bps total (over a day) to 213.6 times the number of devices I have every second of every day. That is 2886 times the amount of data needing processed now.

The other assumption that they are going to move some of the processing local for this, even though they say the devices to not have the horsepower to handle user defined trigger words. This would seem to be squashed as well given the amount of data sent just to process "what time is it".

No comments:

Post a Comment