If you are talking about Android keyboard, is https://github.com/helium314/openboard not a good option?

It can and it will. That is one of the uses of “NPUs” I’m most excited about.

Basically you can run an (potentially open-source) small LLM on the phone using whatever context the keyboard has access to (at a minumim, what you’ve typed so far) and have the keyboard generate the next token(s).

Since this is comptationally intensive the model has to be small and you need dedicated hardware to optimize it, otherwise you would need a 500W GPU like the big players. You can do it for 0.5W locally. Of course, adjust your expectations accordingly.

I don’t know any project doing it right now, but I imagine that Microsoft will integrate in SwiftKey soon, with open source projects to follow.

Four voice typing you could try FUTO which is what i use. Fdroid repo https://app.futo.org/fdroid/repo?fingerprint=39D47869D29CBFCE4691D9F7E6946A7B6D7E6FF4883497E6E675744ECDFA6D6D

For voice typing sayboard may be an option, i’ve never tried it so i’m not sure if it’s good or not though; while for the predictive text i have been looking for a keyboard that can do it as well

While Google isn’t generally good for privacy GBoard actually does this. IIRC they actually completely removed the sync service and your typing history is only kept on-device and Android backup.

However it is a bit of a privacy nightmare otherwise as many of the other features phone home. But last I checked (~4 years ago, worth checking again) the core typing functionality is actually fully offline and private.

So yes, it is possible.

This functionality can be local. I use the Google keyboard with internet access blocked and it works. The only thing missing is the ability to search for emojis typing a word (they are still in the list) and some features that I never used and never understood why they are in a keyboard since it’s not related to typing text, like the gifs.

The only reason I see for a keyboard to need internet connection is to update the dictionary when it’s modified, but it shouldn’t prevent to work with an outdated dictionary.

When I searched for alternatives a few months ago, I couldn’t find anything satisfying.

I’ve been using a version of openboard with sayboard implemented. It doesn’t work perfectly, there’s a good amount of frustration that comes with it, but it works good enough for what I need it to do and I’ll gladly take the trade-off for being less dependent on google.

Hopefully, it continues to get better or a more exact/perfected alternative comes along.

Yes. Very possible. An LLM could possibly be run locally or just sandboxed for only you. In my experience, I guess because there is less training data and fewer iterations, it tends to take longer and result in poorer outputs.

Microsoft could also let you control this but of course they do not want to.

…

I switched to Openboard. After a few months, Its not as good yet as SwiftKey was, but it’s also not sending all my text input to Microsoft.

The primary feature I miss from Swiftkey is the ability to insert a gif easily.