The problem is that our successes have been mainly on human-to-computer speech. Word error rates on the more difficult task of recognizing human-to-human speech are often 20% (or even higher if you don't cheat on the test sets, but let's not get into that).

So why are ASR systems so much worse at understanding people than people are? Many studies have pointed to pronunciation variation as one likely cause of higher human-human error rates. That is, some very clever pervious experiments have shown that when people speak to humans (as opposed to machines) they pronounce words differently. But previous attempts to model this via `pronunciation models' have not had much success. I'm going to discuss a series of analytic experiments trying to understand why our clever ideas, and everybody else's even-more-clever ideas, haven't worked, and what is really causing this pronunciation variation in human lexical production. Hint: people don't like to bore other people, but they don't mind boring their computers. I'll conclude with a discussion of some neat new work we are just beginning which applies some of these ideas to TTS (Text-to-Speech synthesis). Because people also don't like their computers boring.

Back to the Colloquium Page