He Darren,
We should not do in this way, since if decoder broken, then the keyboard hang.
Whatever decoder broken or not, we must handle key event at once.
It's different with virtual keyboard.

I see...I agree we should not block typing.

How about, we still allow users to type (like current behaviour), but we save the characters somewhere in a buffer. When the decoder loads, we send those characters to the language model so that the autocompletion gets the whole word.

I don't know much about how decoders work, so if this is too complicated then we don't need to fix this. I'll just run the automated tests with autocorrection off.