Kaldi is pretty good. Not sure if you can run it on a phone but definitely on a single desktop, purely local processing.

Results depend on the trained model, I think the Tedlium one is alright. And of course quality of input signal - far field/noisy much more prone to errors, that's where the mic array on Echo helps a lot.

Here's a relatively easy way to set it up

https://github.com/alumae/kaldi-gstreamer-server