I agree with the other answer that a noise removal VST (or similar) is probably the most efficient option.
If you want to try something “AI” based anyways, I would suggest demucs, which was originally made to separate music into voice, instruments, …
How is that tiny though?
Considering this is about sending some random data to a server and measuring the speed, that’s quite large. I’ve seen whole computer games that fit in 1/10 of that space.