[geza]

I actually solved this problem back in 2013 with a slightly more advanced technique (taking into account other signals such as motion), see http://up.csail.mit.edu/other-pubs/chi2014-smartsubs.pdf and http://up.csail.mit.edu/other-pubs/gkovacs-meng-thesis.pdf for the algorithm and https://github.com/gkovacs/extract-subtitle for the implementation (in python)