李开复Kaifu Lee-Talking to Machines.doc

BUSINESS TECHNOLOGY; Talking to Machines: Progress Is Speeded Designing a computer-based system that can recognize human speech has long been a cherished goal of artificial intelligence researchers and linguists. After almost three decades of agonizingly slow progress, some researchers believe they are closing in on the ability to create machines that can recognize naturally spoken language from a variety of different speakers. The latest advance comes from Kai-Fu Lee, a computer science graduate student at Carnegie-Mellon University in Pittsburgh, who has developed Sphinx. It is an experimental system capable of recognizing continuous speech based on a vocabulary of almost 1,000 words. In the past, most voice recognition systems have been limited to individual words or words that are separated by pauses. For example, last year the International Business Machines Corporation said it had developed a prototype speech recognition system based on an I.B.M. PC/AT that could recognize 20,000 words when they are separated by pauses. The system is being tested as a dictation tool by several large corporations. The systems biggest shortcoming is that it is cumbersome to use, and it must be trained for each speaker. In contrast, the Sphinx project has achieved an even more important goal. The system can recognize speech from many different speakers without requiring special programming, or what speech recognition engineers call training. This is the first system that has a large vocabulary and attains speaker independence, Mr. Lee said. A number of speech recognition products are available, but they have extremely limited applications. For example, the American Telephone and Telegraph Company and several others are offering a voice dialer for cellular car telephones. The system can recognize numbers that are articulated precisely, and can then automatically dial the phone. The system allows the driver to keep his or her eyes on the road and hands on the wheel. In Japan,