Did you ever find a offline voice assistant, if so do share, there is one that i tried installing on debian (none RPI) but i was unable to get it setup. Since you are looking for a RPI solution here it is
Rhasspy probably offers the biggest flexibility but you will have to build your own integration into openhab over something like nodered or mqtt.
Sepias developer is active here on the forum and i think he build some kind of openhab integration into it. As i have no experience with sepia best ask him yourself here:
Mycroft is not completely offline yet as it works best with google stt but there working on transitioning to mozilla deepspeech and they have a very active community of makers and an openhab skill.
I think those three are your best contenders.
If you got lots of time and want a project you can build your own too.
-There is snowboy as a hotword engine.
-sox and its integrated silence detection to record and do vad
-pocketsphinx python for stt (you will have to build your own domain specific language model to make it work good)
-for nlu you can just start with something like simple regex matching
-tts there is pico2wav
-tie it together with something like nodered
This is kind of valid for SEPIA too, if you want to speak completely free without building your own, small domain language model. I haven’t seen any open-source ASR in 6 years that can match Google unfortunately. The acoustic models of open ASR are getting better and better but Googles TB size language models are crazy
In SEPIA you can switch easily between native and custom STT where native can be Google cloud (Chrome/Android), Google offline (Android, supported on some devices) and Apple cloud (iOS) and custom is the SEPIA STT server available as Docker container using Kaldi open-source ASR. The Kaldi language model can be customized.