If anybody ever opens an S&M themed bakery, the problem will be accentuated.
Now it won't be a surprise.
It might be better then to invest effort in improving language inference heuristics, which would help in all applications (pdf documents, text files) than to try to build in support in each underlying protocol.
This means that if the screenreader picks the wrong language, then the user has to guess the spelling from the way it's pronounced to understand the text.
Though I'll concede that right now on the web it isn't much better; but that's a failing of publishing tools, not the format.
Indeed, so why not let the human writer clarify it?
Why do you prefer a resource intensive guessing algorithm that will never have a 100% accuracy over just having an annotation that takes less than 10 bytes and is trivial to parse?
In my opinion there's no reason to even consider the first option, especially with gemini's focus on simplicity in mind. This "what can another little JS library hurt" attitude is what lead to gemini in the first place.
Add language inference heuristics to cover all kinds of formats and now you've got a dependency. Then some people need different settings so you add config files and parsers/dependencies for those. At some point you update the model and now it can't properly differentiate Mandarin, Cantonese and short Japanese phrases anymore. You start adding cross-platform support and other features, and now some people say it's too slow on their Raspberry Pi Zero setup. Bloggers complain as they have to rearrange some quotes via trial-and-error so the heuristics pick up the correct language. Unfortunately that makes it worse for people running the older version 0.7.
After this rant it should be obvious, but: I prefer just adding a ["en"] or similar and stop worrying about it.
It may be that the language which a quoted foreign-language word is meant to read in was mentioned far back in the text. Human sighted readers will know how to read it, but computers won't yet (and perhaps won't for many years if not decades). However, there are visually impaired people now who need to be guaranteed the same Gemini experience as sighted users.
Obviously humans are a bit better at contextual clues, but.. the information has to be there for a human user to make the distinction.
Not always, but often enough that it's an informal standard.