I would like to add that HeliBoard (keyboard) in combination with Whisper provide a fully offline experience. Whisper needs to download the Whisper models from Hugging Face (~ 435 MB). This is the only use of the internet permission. Then voice recognition works completely offline (albeit a bit bumpy and not for all languages). There is a fast and slow model. In any case, it would only work for more recent devices.
It would be interesting to see if the murena / e/OS folks could team up with Mistral (https://mistral.ai/, based in Paris, France, more inclined to open source). The hay build the Mistral 3.1 model, which works quite well. To the best of my knowledge, there is no plan to support STT.