Let me give you the first informed answer in here.
1. Google and Facebook are not listening to you, and in general normally also Amazon isnt.
2 What is happening has to be separated in two categories.
a. False positives. People misinterpreting stuff - happens all the time, dont feel bad about it.
b. Data congregation (correlating data). Even though Google, Facebook and Amazon arent listening to you - other app vendors are. There are apps out there that track audio on your surroundings to f.e. identify what TV programs you are watching, to correlate that with your twitter stream, and so on and so forth.
Those other programs are selling their data profiles, those get bought by aggregators (the 7-10 businesses you've never heard of, but that have bigger personalized collections than FB), which then in return also facebook and advertisers use - to - "personalize your ads". (Or target you for psychological profiling on political ads prior to political votes. Its a thing..
).
Now - two things on that, first there has been a recent set of articles on this stuff (third party apps listening to live audio, selling keyword matching), I have to read those, I havent so far - will link to them, once I've sourced them again. Second, its still somewhat hard "processing intensive" to do that. (Might be using clickworkers to do it, who knows...
) So I'm not sure how viable this has become yet.
Not everyone has the neural networks to do real time voice recognition, and usually (if you've "talked" to an Alexa its still quite poor, if you dont limit yourself to a "preset range of words" - think "automatic youtube subtitles are crap").
//
Why we know that they arent doing it. Two aspects again, one - we can analyze "normal" data traffic (if its encrypted its harder, but man in the middle sometimes is possible, and we can look at data volumes). So we would in general see, if those big companies do "full takes" on voice data. They arent.
Second is battery consumption - your smartphones would be empty in four hours, if someone would be always doing a fulltake on voice data, and uploading that. (Alexa is a powered device - yay!
)
How those (useless btw) smartypants voice devices work, is via keyword recognition. So they are "listening all the time" - but only for a set (prestored pattern). Think of it as the computer only knowing one word at all. Thats rather easy and can be done on cheap arm based hardware. Once it thinks the keyword has been uttered, it then records the next lets say minute of the interaction, and uploads the entire voice sample to cloud processing, where more processing intensive proprietary analysis is done.
Those uploaded bits are all stored for the next 100 years (all 30.000 of them you made over the past 5 years), if you dont object to the collection (if you can - with GDPR you should be able to, please double check). Thats a thing.
//
Special "targeting" cases.
Can the police.... (get access to my Smart TV with camera and mic, or my Alexa, or my smartphone...)
Depends on which police..
In theory they could, and the companies dont have to tell you thanks to recent US regulations (terror, terror, be very afraid), but that would be usually at least somwhat targeted, so the "police" would have to have an interest in you somehow. (For example, because your best friend is on a terrorist watch list..
) "Normal" police work usually doesnt entail those "levels of access". In recent months they have been starting to ask Amazon for voice recording data, but since Amazon (and others) only upload once a keyword is "identified" - the data they could get out of that is spotty.
Can they "link themselves in" in real time - if they have the required legal blank to do so. Maybe. Not even probably right now. Because there were no recorded cases. Currently you'd be second guessing state agencies behavior at this point.
With cellphones (not even smart, just normal is enough...
) we know, that they can do that. Its baked into the cellular network stuff. The way it works there is, that they send whats called a "silent sms" (an sms that doesnt show up on your phone once its recieved), and your phone goes "blub" (non audibly..
) and starts turning on the microphone, broadcasting your surroundings as a telephone call. Similar (not same) with location data (triangulation, cell tower stuff).
Thats the thing why "in the movies" people put their phones in fridges, or Jason Bourne slams his girlfriends phones on the street.
//
So in general those big three arent listening to you at all times, but smaller app vendors might - although it shouldnt be very commercially viable at this point.
(Its relatively simple to find out which TV progrem you are watching through voice recording, its harder to identify if you have an interest in lets say Tesla or the Bahamas, by listening to your daily conversations. Its much easier, to wait for "intent" on your part (like you doing a google search, or writing/recieving an email, some vendor scans for keywords) in those "open end - could be anything" queries).)
But now, let me search the last bunch of articles on third party app vendors doing full voice takes, and selling that data.
(Correlation, then means, that that data gets matched to the "profile" some vendor has about you (any unique identifier will do it could be email address, could be phone number, could be....) - thats happening all the time (several people are collecting, it all ends up under one profile - for your person). Thats a business.)
--------------------- MERGED ---------------------------
Google isn't showing any rencent "usable" articles on this "debate", so those two have to do:
https://www.wired.com/story/facebooks-listening-smartphone-microphone/
https://www.vox.com/the-goods/2018/...ok-microphone-tapping-recording-instagram-ads
Here is an IDIOT vice guy talking about "smart devices could have thousands of voice triggers":
https://www.vice.com/en_uk/article/wjbzzy/your-phone-is-listening-and-its-not-paranoia
Yes, they could have - but they probably havent. With false positives that would mean that they are constantly sending voice data and they arent. This also would mean that actual keyword phrase identification accuracy would be lower and... Essentially the entire article is without any basis and bullsh*t. Now could they have lets say 10 "secret" activation phrases, that maybe get changed over time? That would be more sensible, and harder to identify. But it would be malice nevertheless, and the companies are denying that, ... and it still wouldnt be very economical to do so.
If theysomewhat know - that you are sitting down to watch TV (you opening up a TV guide app f.e.) voice id'ing what you are watching makes sense. Just listening to you all day to find out if you need a t-shirt, or what your favourite jeans brand is - doesnt.
Thats easier identified, by finding out which social group you belong to (because you liked favourite singer, or even brand), and then giving you the top 3 picks of advertising for that group and age bracket. No one needs to "listen to you" for that.
Facebook and co have thousands of profiling identifiers for you. They even show you a fraction of them if you click on the corner of some ad. Those are much easier to set up, serve much better to predict your every behavior, wants and needs, without having to listen to you babbling about random stuff every day..