http://www.antoniogarciamartinez.com/about-antonio-garcia-ma...
> To make it happen, Facebook would need to record everything your phone hears while it's on. This is functionally equivalent to an always-on phone call from you to Facebook. Your average voice-over-internet call takes something like 24kbps one way, which amounts to about 3 kBs of data per second. Assume you've got your phone on half the day, that's about 130 MBs per day, per user. There are around 150 million daily active users in the US, so that's about 20 petabytes per day, just in the US.
No, all they need to do is use the client device to perform speech-to-text and send the tiny amount of resulting data.
I don't think Facebook records people's conversations, but I also don't buy this article's other arguments that 1) conversation data is not useful to advertisers and 2) Facebook would not at least make a semi-successful attempt at analyzing natural language data with their existing NLP tools.
> Because it has no specific trigger word for Facebook, your phone would need to listen for every targetable keyword. That means the speech-to-text translation code could only run on your phone itself, a taxing demand even for the beefy cloud servers that usually handle those tasks.
What app currently does speech-to-text translation with any usable accuracy on the device? Even Siri and dedicated-devices such as Alexa don't seem to attempt anything beyond trigger-word recognition locally.
On an iPhone, Apple has the full run of all device capabilities, and can even ship custom chips, but they still stream audio back to a server to process what people say to Siri.
Speech to text is very hard even when the speaker is purposefully speaking to the device. It would be way harder to try to extract meaning out of ambient audio.
random sampling and building a profile overtime that can be used to enhance the sampling is a very easy method to snoop at times that gives you good results.
PS I don’t think “they” are listening and use other highly effective means to target ads. But I do think arguing such targeting is only possible in the cloud, and only after recording all the sound in your environment: the air conditioner, traffic, kids playing, etc. and let’s not forget ALL the silence in between, is intellectually dishonest. Especially when we all know that any always-listening device processes the signal locally, scanning for trigger words.
I think he smudges the issue by conflating FB's theoretical ingestion of such audio with its current data storage capacity and ingestion. Presumably, raw audio could be transmitted and processed without being stored.
But does the state of the art language recognition allow for broad on-device translation beyond a subset of trigger words, i.e. "Alexa", "Siri", "OK Google"?
You're talking about two different processes, recognition and comprehension.
Recognition on device is quite feasible and could be expanded from the trigger phrase to several dozen or perhaps a couple hundred keywords without burning up the device. The iPhone 4 did this on-device, pre-Siri, for voice dialing. Many cheapo car stereos do this too. And it is what the intelligent speakers and phones already do without cloud interaction, scanning for that trigger word.
Comprehension on the other hand, deciding what exactly the speaker is asking for, requires compute-intensive NLP and AI processing, and yes, that's going to the cloud.
Back to the specious argument by the article's author, to create an audio ad targeter doesn't require constant streaming to the cloud. For starters, you don't need to send silence or background noise. You don't need to send every word spoken either. You can screen for keywords in the trigger list, then send what is stored on-device locally before and after the trigger and forward just that.
Again, I don't think Facebook is listening to target you. Maybe if some CIA front corporation buys them up... I do however think the author's argument, the very first argument he made (and thus my labeling it as the foundation), is a load of malarkey.
No. You think Apple, Amazon, and Google ship all those audio packets back to servers for fun?
To believe that Facebook is processing ambient audio on a mobile phone, you have to believe that FB is doing things on a phone that even the phone's creators (who have no sandboxes and can access custom chips) cannot.
My point was that technicality isn't what will get in the way of listening vs not listening. He's backing his claim by saying it's technically impossible but that's just weak defense.
There are plenty of ways to do it like sampling / increase sampling when other fb friends are around or maybe they just listen randomly and in the end they’ll have a well built profile about you.
It’s not like it needs to be listening all the time or respond anytime a keyword is said.