You'd all but certainly be better going in the other direction.

Text is a more constrained state space than speech/audio.