I could see there being use for something similar to this targeted towards children for expanding their vocabulary.

Similarly, how neat would it be to have a program that does this for object in video, and testing the viewer on the name of the object? Perhaps to avoid error as a result of imperfections in the neural net, it would require 2 parties, the learners and the teachers. The teachers would go through a video ahead of time with the video, making sure there are no errors, then that session would be saved and the teacher could play it back for the learners.