That idea crossed my mind even seeing how many companies failed even during boom times.

Service could be powered by some speech recognition software (maybe more than one) which initially would give insane percentage of incorrectly recognized text. Community could fix those errors (weak link in my plan:)). Afterwards you feed data back to software and train it so next time it would work better.

Training would require marking which person talks when in episode to tune algorithms for each of them. Maybe group them by regional dialects.

Oh, another tiny small thingie - not clear how to monetize that.