I’d like to see the source that says Dragon’s WER in the 90s was 3-5%. I used Dragon in the 2000s and it just wasn’t comparable to the current state of the art.

https://dragon-medical-transcription.com/history_speech_recognition.html, for example. a lot of adverts and awards were given to it (admittedly awards like PC Mag that were probably paid advertising... but that's why I went with Open AI's assessment on whisper at 2%.) Dragon was boasting 99% accuracy after (admittedly months) of training; and it frequently reached it. there were some gotchas in that- the months-long training was a big one. The other was that you frequently had to slow down and be careful to enunciate that you don't have to do with modern systems (including the MS versions of Dragon- they bought it out at some point)

whisper.cpp is an opensource implementation, although I’m not certain exactly how open.

It's on the MIT license, if that helps. I take issue with anything OpenAI is involved in. for oh-so-many reasons.