Android has about two billion users - there's an extremely strong incentive for large, ad-driven companies (like Facebook) to segment Android off and take advantage of every API possible to gather data with - and they certainly have the resources.
> Or do you imagine the hypothetical speech-to-text modelling is happening on your phone?
Yes. Speech recognition is a relatively easy problem for the purpose of the advertiser, because they don't need anything approaching perfect accuracy - unlike tools like, say, Siri, which need a very high accuracy rate. It's well-known that you can sacrifice accuracy for performance/power in many algorithms, and voice recognition has been the focus of intense study, with many low-power algorithms being developed. For instance, here's a design (it does contain a custom chip, although one tested at a large process node and containing functionality similar to a GPU) that consumes 8mW of power on a >100k word dictionary: https://groups.csail.mit.edu/sls/publications/2018/Price_IEE...
> For an example of a similar battery draw: try opening Shazam or another "audio fingerprinting" app
Those are completely different tasks. Audio fingerprinting is not remotely comparable to low-accuracy voice transcription - otherwise Siri and Google Assistant would drain your battery constantly at the same rate.
It's pretty clear that your claims of "superstition" are unfounded speculation that is inconsistent with the abilities of modern technology.