Staff writer
Scientists have long dreamed of developing electronic products that can be operated by verbal commands, and that fantasy is becoming a reality, according to one of the world's leading experts in speech recognition systems.
"I think we are at the level where clearly these products are available," said Alexander Waibel, principal research computer scientist at Pittsburgh's Carnegie Mellon University and professor at the University of Karlsruhe, Germany.
His international research team recently developed a state-of-the-art speech recognition system called JANUS, which has received top marks in performance tests by both the U.S. and German governments in the past few years.
It can recognize speaker-independent, spontaneous telephone conversations -- the most difficult task in speech recognition research -- and last year achieved the lowest error rate of any system -- 26.7 percent.
High dictation performance, such as comprehension of newspaper articles read aloud, is another strong point. It has an error rate of less than 7.8 percent in English and 8 percent in Japanese, he said. "The technology we've developed focuses on spontaneous speech (so it can handle) sloppy speech. That is a strong feature of our technology," Waibel said during a recent stop in Tokyo to give a presentation to companies interested in using his system with their products.
With an effort to provide greater access to speakers of different languages, Waibel and his research team now offer the system in 17 languages.
Using JANUS technology, Pittsburgh-based Interactive System Inc. -- where Waibel serves as chief technical adviser -- has created a software kit for developing Japanese voice-recognition applications with Tokyo-based venture company Advanced Media Inc.
Telephone companies such as AT&T in the United States already use voice recognition systems, but a system that recognizes spontaneous speech from various users is not yet commercially available, according to Waibel.