I actually worked on this exact problem as an intern job at our university. We used a huge corpus of communication (for example, we had access to all the emails every sent internally at Enron).

We used this as the basis to train a speech-to-text engine by automatically correcting likely-wrong interpretations. "I go loo school" would be corrected to "I go to school", for example. It worked remarkably well.

The basis of all these subtitles can be used, but there are far bigger (and better?) collections of data to be used to train these machine learning engines.