Automatic phonetic transcription of speech: is it still a difficult problem?

Question

A friend was asking me why there aren't any programs that can analyze a user's pronunciation of words and then give feedback/suggestions.

Indeed, given an accurate phonetic transcription of a predetermined set of words, determining if a user's pronunciation of each word is standard/correct — or, say, determining someone's regional accent — should be trivial.

However, my sense was that Automatic Phonetic Transcription (APT) is not nearly accurate enough to analyze a user's pronunciation in that way — though my knowledge is about 10 years out of date.

Still, it made me wonder: even if APT of spontaneous speech does have a low accuracy — in a context where each expected word (and even possible alternate pronunciations) is known ahead of time, would accuracy conceivably high enough to give a user feedback on pronunciation?

Edit: arjan pointed out that Rosetta Stone claims to be able to give pronunciation feedback; how does such a system work (and how well)?

Answer 1

True automatic transcription is still nonexistent, unfortunately. The closest thing that I know of is this thing called Blitzscribe, which seems to be reasonably accurate (88–94%). This is what was/is used in the Human Speechome Project, which is also very interesting.

Here are some sources for further reading: The Study on the success of Blitzscribe and Ted Talk by Deb Roy, the guy who started the Human Speechome Project