(1) What they are speaking - IN their script (this is called SLS: Same-Language-Subtitling and has been quite successful for literacy programs for one's own language)
(2) What they are speaking - IN my script
(3) What they MEAN - IN my language (i.e. translation)
Data for (1) is easy to find. (2) could theoretically be generated from (1). It's (3) which requires a lot of deliberate manual effort.