The claim of the article is that it would be technically infeasible - and they go into how difficult it is to interpret context. However simple keyword matching would be more than enough.
Facebook has a lot of compute resources, but they wouldn't have to use it. Your smartphone is more than fast enough to do simple speech recognition. The accuracy rate wouldn't have to be that high - you won't get mad if you see an ad for a misheard keyword.
I was thinking the same thing. Also not claiming to have any real idea of as to what they are actually doing but certainly having access to everyone's phone gives them a ton of distributed computing power for free.
>I was thinking the same thing. Also not claiming to have any real idea of as to what they are actually doing but certainly having access to everyone's phone gives them a ton of distributed computing power for free.
Only if by "everyone" you mean people foolish and vapid enough to use Facebook and give them access to your phone.
Radioshack sold a voice recognition chip in the 80s[1] that was a simple 8-bit microcontroller. If you are willing to slip on the accuracy and false positives you can do recognition with very little computation.
the downside was power usage. Motorola made one that was power efficient, used by nokia in the 90s and its pretty much the same chip in google's phone line today (just even more power efficient).
The Google Now only listens for the trigger phrase when idle - which is done all locally, without needing to talk to the servers.
It has a battery impact but much less than sending all the voice data continuously to a server somewhere. The biggest battery killer would be the wifi or 3G transmitting non-stop in that case.
It wouldn't have to transmit non-stop -- it could do some parsing/cleanup locally, then queue it up and upload it periodically with other, expected FB traffic.
Ad is easy. You don't have to understand context. Just listen for a thousand or so keywords related to products that are paying you. Then if detection happens apply some rudimentary sentiment analysis on the surrounding phrase and that's all you will ever need.
if you have a couple millions for me to start a small team we can offer this as a service next month or two.
Facebook has a lot of compute resources, but they wouldn't have to use it. Your smartphone is more than fast enough to do simple speech recognition. The accuracy rate wouldn't have to be that high - you won't get mad if you see an ad for a misheard keyword.