I think exposure and actively speaking are the wrong categories - your learn a language by using it to participate in a common experience with another human being. For this, you both have to use the same protocol, and at the beginning, you have to figure out this protocol using a simpler fallback feedback protocol. This is something I am currently realizing as a father of a one-year old. This involves of course exposure and speaking, but I strongly doubt that you could learn a language just by listening to it and re-producing the words. I expect this would only result in text/speech resembling GPT-2.
I also think that you can have a common experience without actively producing text or speech, for example by watching a movie / TV show were you want to follow the story and participate in the lives of the actors. In fact, I know many people who effortlessly learned a second language by watching foreign movies and TV shows for years.
I think the way that we are able to process speech at that speed is that we essentially guess what words are coming next, which is really only possible if you’re able to construct the sentence yourself. If you stop a sentence at a random point a native speaker will be able to predict the kind of words that will come next, if not the exact words. Another bit of evidence for this theory is that if someone says something very unexpected and out of context we often won’t understand what they’ve said.
Now it’s possible that you could achieve this ability only through input and not through production but I believe that it will be much less efficient.