Yeah you can extract them. I've not tested it yet but you give it the narrator device and translator library files from a WB 1.2 disk, it'll create IFF files that are loaded instead of the free ones that are built in (they were used to make fixtures for byte-identical testing). But unfortunately will only give the Workbench 1.2 sound at best. From 2.1 it used a different system that I haven't even looked at yet. @Kalamatee has a decent start on a modern alternative using machine learning though, so we could eventually have 3 back-ends. My first attempt at it is here: github.com/bitplane/AROS/tree/speech.device but probably not worth looking at it yet as I haven't really, I vibed pretty hard and need to properly rip it apart in review before putting any faith in it! Decent proof of concept though