It's purely business reasons. Also in Chrome, the STT runs in the Google cloud anyway.

And just FYI I'm working on turning my state of the art speech recognition paper (TEVR, token entropy variance reduction) into a developer-friendly binary which will run offline on your GPU and, hence, be fully private, and offer a scripting API on localhost. Testing WER on LibriSpeech clean is 2.3, so slightly worse than an offline wav2vec2 1B, but years ahead of Kaldi, Vosk, coqui and the usual streaming cloud services.

I estimate it'll be 2 more months until I can post Linux binaries. Won't be open source, though.