Bilkent ACM SIGART (Special Interest Group on Artificial Intelligence)seminar...
Abstract: Speech recognition -- the problem of taking audio data and converting it into typed text -- is an artificial intelligence problem which is not so difficult as to be entirely out of reach; but still hard enough to make us have to think. State-of-the-art systems have error rates for telephone speech by unknown speakers (speakers on which they have not been specifically trained) of about 30%, and the algorithms take hundreds of times longer than real-time.
I will give a brief introduction to current speech recognition techniques, and suggest why I think the progress in the field is so slow-- namely, the search for a "neat" solution. Instead, I suggest that there may not be a neat solution to the problem of speech recognition, and that we need to acknowledge that the solution may be messy (as, indeed, the biological solution is messy) and focus on ways to find that solution. This may involve sharing code as well as ideas.
Thursday 30th November, 2000
at 17:40 in EA502 (Eng. Building)