After a vocabulary item is answered there should be some options for:
“My answer was right and Wanikani didn’t accept it.”
“My answer was wrong and Wanikani accepted it.”
Then you could record those answers and if people are typing in a correct answer than Wanikani isn’t taking it would start to learn them.
"My answer was right and Wannikani didn’t accept it."
After a couple people do that it starts accepting “quiet voice” as an answer.
A full machine learning implementation could be used, but even simple implementations would likely do.