Here's an excerpt from an article about the future of speach recognition .
The value of speech products in telephony systems -- including equipment, software, services and license fees -- will exceed $11.6 billion by 2001, most of it in limited-vocabulary speech-recognition systems, estimated speech-market analyst William Meisel of TMA Associates (Tarzana, Calif.). Some $285 million worth will be in speaker-verification systems and $308 million in text-to-speech synthesis.
The speech-recognition problem, in fact, is more than just a matter of recognizing words. The comments of researchers like Levinson, Wilpon and Picone suggest that it may be fruitless to ask the computer to recognize every word and proceed from there into an understanding of what was intended. Rather, dialog systems would have the computer recognize as many words as it can, but formulate its responses on the basis of key words in a sentence. In effect, it looks for the words that convey context-sensitive meaning.
Because of the discontinuities in speech, said Fred Juang of Lucent Technologies (Murray Hill, N.J.), who also is editor-in-chief of "The IEEE Transactions on Audio and Speech Processing," it does not make sense to try to convert all spoken sounds into words. Rather, the applications that Lucent will concentrate on will utilize what he calls "key-word-spotting technology." Extensive research and field trials conducted by Lucent's Bell Laboratories revealed that basic accuracy of a command-and-control system increased dramatically when the machinery used only key words in its recognition search tree. The results of this research won't lead to a conversation with HAL, joked Chin Lee, who heads dialog systems research at Bell Labs, but it will make the man-machine interface more acceptable.
techweb.cmp.com |
