You speak a sentence (or more). The computer transcribes it to text, but also displays in parallel with the transcription a probability diagram indicating regions of high uncertainty. (Google Voice uses text shading.) You click on one of the incorrect regions, and it displays possible corrections in that region.
The problem being addressed is, if transcriptions are given in order of decreasing likelihood of the whole sentence, there may be many choices not affecting the incorrect region.
There is a problem that probabilities of words are linked across the sentence: The dock chased a cat, and then that same dock chased a squirrel. You click on "dock" correcting it to "dog", but then the probability of the second instance of "dock" changes (or should).
No comments :
Post a Comment