02/23/19 ONLINE ORIGINAL: IT WORKS IN THE LAB, BUT WHAT ABOUT IN THE MARKET?

Return to main story

IT WORKS IN THE LAB, BUT WHAT ABOUT IN THE MARKET?

Two years ago, Steve J. Young, a Cambridge University professor of information engineering, co-founded a speech-recognition company, Entropic Cambridge Research Laboratory, with his colleague Philip Woodland. The company is based on university-developed software that helps computers understand human speech. In late December, Young merged his Cambridge company with its Washington-based cross-licensing partner, Entropic Research Laboratory Inc., into a company now simply known as Entropic. The merged company currently employs about 50 people, with small operations in Washington, D.C., and California, and its headquarters in Cambridge, England. Young hopes the merger will help Entropic expand more quickly and attract needed investment -- particularly from the U.S.
By one measure, Entropic's technology is world-beating. For five years running, the U.S. Defense Advanced Research Projects Agency (DARPA) and the National Institute of Standards & Technology have ranked it first against such larger rivals as MIT and SRI in an annual evaluation. Yet, Entropic remains commercially overshadowed by companies such as IBM and Belgium's Lernout & Hauspie. Indeed, Entropic's 1997 revenues are expected to be a puny $4.5 million. Customers include Bell Laboratories, the Defense Dept., and Harvard University. But Entropic may get some help on the commercial front soon. It could be one of the first investments made by Amadeus I, Cambridge entrepreneur Hermann Hauser's new $65 million venture-capital fund. "We very clearly need to expand and grow to meet the challenge from Lernout & Hauspie and others," says Young.
PSYCHO-TRANSLATOR. One place to see Entropic's speech-recognition products in action is at the company's modest, modern offices in the center of Cambridge. I came away bemused, but not very impressed. The first product that Entropic CEO Brian Corbett demonstrated was the company's "grapHvite" recognition prototyping system, a program for software developers. Talking into a headset, I told the computer "Give me four tickets for this evening's show of Psycho, please." (This was a program designed for movie theaters to sell tickets over the telephone.) Instead, the computer translated my request as, "I'd like four tickets to see Mission Impossible, please." The second time around was even worse: "I'd like two tickets, please, to buy....to see Boogie Nights" got translated into "I'd like 10 seats to watch Silence of the Lambs." In fairness, of course, Boogie Nights wasn't programmed into the computer's vocabulary, so it substituted what it thought was the closest equivalent.
But whatever the excuse, there is clearly a gap between the technical superiority of the program developed by Cambridge University researchers, as judged by DARPA, and at least a superficial test of the commercial program based on that technology sold by Entropic. Corbett's explanation was that the variance and numbers of mistakes would be reduced by what the software developers then did with the program -- such as adding vocabulary, ways of asking the same question, etc. Also, grapHvite is a program aimed more at interactive dialogue and less concerned with exact dictation as long as the basic facts are right, Corbett says. Yet my impression apparently wasn't shared by a recent industry trade show, which according to Entropic, voted grapHvite winner of the 1997 best of show award for speech-application developers tools. And Corbett did show me a transcription product that took down verbatim what someone says into it. This program made far fewer mistakes than grapHvite did.
Young says Entropic's biggest challenge isn't technical, but rather to "package speech-recognition and speech-output technology in a way that can be used to build systems that can communicate naturally with people." In turn, the business challenge, as he sees it, "is to convince major companies that we have the best technology -- and not to buy from the safe option of very large companies." He adds that while he's confident Entropic can deliver the technology, "the main challenge is breaking into world markets. Convincing companies to buy British technology is a bit harder than convincing them to buy American technology." Entropic's strategy is to focus on providing natural interfaces to general users of computer information systems, such as telephone answering systems and call systems that will involve talking with a machine. More generally, the company's focus is on continuous speech. He says any large company with a customer service operation is a potential market.
THAT'S SOME TRICK. Entropic's technology is based on the Markov model. Essentially, it's a software program that involves calculating the statistical likelihood of certain words following others, built upon 1) a model of basic sounds, 2) a dictionary of words, 3) a language model of one-word sequences following others. To hear Young tell it, all speech-recognition systems basically work this way. The trick is to try to make the acoustics models as accurate as possible and find the sequence of sounds and words that are the best match to the input. To do that involves creating clever algorithms that search all the possible sequences.
Why the difference in results between the DARPA evaluations -- which have repeatedly ranked Cambridge No. 1 -- and Entropic's commercial software? They have different parameters, Young says. For the DARPA competition, the various teams can take up to three weeks to recognize a couple of hours of speech. The evaluation measures how few errors the programs can make if constraints are few. In contrast, the goal of the commercial software is to get the computer to recognize speech as quickly as possible, while keeping the software as lean as possible in terms of memory requirements. Likewise, the commercial software has to be more robustly engineered and backed up with technical support. Young compares the Cambridge technology to a car. "Everyone knows how the combustion engine works." he explains. "It's just that some people can make it run faster."
By Julia Flynn in Cambridge

Return to main story