Would be cool if you could run these speech models in tandem, and compute a new error-rate for the consensus of them.