Siri can't even tell if I'm saying "apartment lights" or "Obama lights" when I'm talking directly to it. And if it's not even hard enough to transcribe all the ways all the people across the globe might pronounce words in all of the languages they speak, there's the problem of deriving who said it (was it you or was it the TV?) and deriving what was the expressed sentiment.

It's so much cheaper and technologically easier to parse your browsing history and parsing things you enter in as text than it is to parse voice. Especially when you don't own the platform your software is running on (what would need to happen for Apple to allow Facebook 24/7 access to the iPhone microphone)?

Also, Google/Apple/Amazon/Facebook don't have to respect our privacy, but they do need to respect privacy laws like the GDPR, so they can't actually tap your voice without your consent unless they want to risk actual billion dollar fines.