Do you really think they're running constant speech recognition on all phones, apparently unnoticed so far, just to recommend YouTube videos?
It'd be much, much easier for apps like FB messenger to transmit sparse samples of low bitrate audio to a server for processing there. 1 hr/day @ 8kbps ~ 110 MB/month. As a matter of fact, with "deep learning" and whatall, it doesn't seem like the audio would even need to be intelligible to human ears for it to still be useful in some marketing algorithms or to generate guesses of interests correlating to the audio. Those interests might be presented as an ad, search suggestion, etc. Why would google use something like this for search recommendations? Because it's training data, that's why!
What I'm arguing is that even though conventional "speech recognition" is processor intensive and unlikely happening on a mobile client side, it's just not that far-fetched that small snips of low bitrate conversation data might be used to tune marketing prediction models on the server-side.
In training general purpose models, huge amounts of data could be generated by only a little bit of data/user. Where would this sort of thing cross a privacy line? I'm not sure. Personally, it kinda creeps me out regardless.
Seriously, listen to these 8kbps opus samples: http://www.opus-codec.org/examples/ . If they're intelligible to human ears, I can't imagine what some clever learning algorithms could do with that and much much less.
Here's a random anecdote about how much data FB messenger can use: https://www.reddit.com/r/androidapps/comments/49yhlh/faceboo...
The fact that it's possible to eat a child or make a puppet does little to make this kind of thing much more realistic.
The much simpler answer is that you are more predictable than you think, and likely more influenced by things Google can watch (clicks, ads) than you think. You also don't spot when it doesn't recommend something so accurate so it could even just be chance (or at least doesn't mean they're good at it)
Facebook messenger and google voice search are my number one suspects. But moreso, all it takes is a single person in a room full of phones to have installed an app that's using the microphone unethically, and suddenly everyone's conversation is being turned into marketing/spy data. Anybody who's connected somehow to that location is going to be statistically associated ("implicated") with this data.