I had a huge backlog of tech videos, so I wrote me this (also to play a bit with Haskell, the base idea can be replicated easily in any language though): https://github.com/rberenguel/glancer
I’m considering a master keyword list to index against any text that comes in.
Any addition that is on the "view" layer (the generated HTML) is very easy to add, just needs to go into the template file, at some point I might tweak that area but currently have no outstanding idea/requirement. The rest (i.e. the bulk of the code) is just a very bare-bones parser for captions that should be pretty stable and need no additions (crossing fingers here).
I’ve seen vosk used on device and it’s decently quick too on a recent Apple chip.