Subtitle generation prototype

This adds a pretty major step towards detecting background noise.
If we know when someone talks we assume something interesting happens.

To do that we need to know when someone is talking, and to show I can do that I generate subtitles.
They're out of sync at the moment but the change is so big I wanted to get it in, because I'm working on the frontend right now.

I also need to make a debian binary again, but later.