@fireborn @alexhall Haven’t tested this yet, but this sounds great! Especially because we might be able to run it in work profiles (the TalkBack Braille IME can’t run there). In place of pure beam search against a dictionary, have you considered something like GPT-2 or another very small (by modern standards) language model? You’d probably get near-realtime inference on most mobile chips. Also the source code doesn’t seem to be in that repo
bill@mastodon.n6.io
@bill@mastodon.n6.io