It will be very interesting to see how this approach will work with dictated speech.

Also let's not forget that the word-level error rate can be reduced by using statistical data about words sequences.