If we currently are at 1/8 word error rate with the speech recognition software, that implies that the error rate of the speech recognition software combined with the translation software will be slightly higher. As indicated in this presentation, however, there is still a potential to reduce the error rate if a larger amount of data is used to establish the statistical combinations.
As with any innovation though, it is still worth having reservations of some kind. Deep down, I feel like this innovation would only bridge communication barriers at superficial levels. The nature of human speech is extremely complex and is not limited to just spoken words. Additionally, spoken words only make sense within certain cultural contexts. It therefore would require some sociolinguistic competence on the part of this innovation to be effective. And that is something we are still miles away from!