Something else: if I had decent amount of music in a game I'd probably go with a MIDI soundfont.

These can be pretty hefty too, though there is the Arachno soundfont which is ~71MiB for a decent selection of instruments. Which I'd trim down to only the ones I actually use (especially since it directs you to credit soundbank creator info individually). Though I think many of the instruments probably depend on dynamics (like note velocity from a MIDI keyboard) to sound good.

Similar for synthesized sounds (non-AI), like sfxr variants, though those might fall on the retro side. And that won't offer complex stuff, voices maybe could use something like DECtalk but I can't imagine getting that up-and-running and also getting it to be multi-platform (I don't like how most modern TtS options sound, MacInTalk/PlainTalk was also iconic though I've never had the hardware EDIT: I just noticed this so maybe it's not impossible. Maybe even use the available WASM as a library via WASI?).