[rpdillon]

FUTO is mediocre with swiping and predictions in my experience, but the 70M parameter voice model is stunningly good at 30-second voice to text. It has completely changed how I think about using my phone to draft prose - my first drafts are now often from my phone, snippets collected in moments when it occurs to me on-the-fly. It's been a really significant shift in the utility of my phone, and because it's installable through F-Droid, I have it on my Amazon tablet, and Boox reader. It's worked really well across all of them.

[Ey7NFZ3P0nzAe]

They are aware of that, that's why they recently decided to create their own swiping dataset. As of today they have not released the updated swiping functions after training it again but it's in the pipelines and I'm really excited for it.

[stavros]

Have you tried it recently? Their predictions/corrections were great for me.

[bogwog]

Ive been using it exclusively for months, and while I do like it, I ended up turning off predictions because of how bad they were. The swipe typing feature is also pretty bad, which is my biggest gripe with it since I love typing like that.

[Groxx]

Yeah, it's choosing extremely esoteric words for me, and clearly missing the beginning of many swipes (despite drawing the trail correctly). Almost unusably bad swiping imo.

E.g. "so" with a nearly prefect straight line clearly drawn within both keys -> it chose "SSSI". What even is that? Or "drawn" there chose "Den", which is apparently a proper noun because ("cause") I didn't ("Ivy") hit shift.