Janet Baker’s Optimism

With seven versions of its DragonDicate 30k currently underway, nothing should stop Dragon Systems from consolidating its position as the preeminent supplier of speech recognition systems. The company is now going in other directions, too, with a prototype Japanese–English mt system now under development (See  sidebar ).

The first thing that Janet Baker, president of Dragon Systems, offers her visitor is a quick tour of the new quarters in which Dragon Systems has recently settled itself. There is a lot of empty office space around Route 128, the renown Silicon Belt of Massachusetts, she explains. In particular, the mini-computer companies in the area have had a rough time in recent years. You can choose whatever kind of space you like, she says. After a thorough search, Baker and her company decided to move into this splendid old 19th Century factory building in Newton, a suburb of Boston. Originally a rope factory, the building is long and narrow, with huge, ceiling-high windows and immense wooden beams.
    The building has only been partially subdivided into office space; workmen are still finishing the upper floors. “There’s plenty of room here for expansion,” says Baker confidently. Dragon, one circumspectly conjectures, has not suffered from the ongoing malaise in the computer industry nor the more general recession in this part of America? “Are you kidding?” responds Baker vigorously, “Our turnover has quadrupled in the past twenty months. We are expecting a real explosion when this technology catches on.”
    Janet Baker and her husband Jim Baker founded Dragon Systems in 1982. Previous to that, the couple had worked at ibm in the 1970s before joining Exxon as part of that big oil company’s ill-fated attempt to enter the office automation market. The first Dragon product to see the world was a speech recognition system for the Apple II. Originally, it had a vocabulary of sixteen words; eventually this grew to thirty thousand over time. Its speed, says Baker, was quite acceptable: users could achieve from fifteen to sixty words per minute. “We have always believed that it’s important to be able to offer users a working system, even if it isn’t perfect,” says Janet Baker. The Apple II system formed the basis for the company’s subsequent developments, financing its further efforts. Despite its substantial investments in R&D, Dragon has remained privately owned and has continued to expand without venture capital.

Friends and rivals
Currently, Dragon’s income is roughly divided in half between direct sales of its systems and licensing agreements with third parties. “We have always believed in forming strategic alliances,” explains Janet Baker. Dragon collaborates with major players, such as ibm, dec, and Xerox, smaller companies, such as Lanier and Articulate Systems, which specialize in a given field, and international partners, such as Lernout & Hauspie (Belgium) and Aptech (uk), for the foreign language versions currently under development. “Unfortunately, people have demonstrated speech systems based on our technology without always acknowledging it,” she adds ruefully. Baker estimates that Dragon has a five-to-seven year lead over its rivals in the speech rec arena, which includes ibm, the Bell Labs, Kurzweil Applied Intelligence.
    DragonDicate 30k, the company’s flagship product, was introduced in 1989. This innovative product is the first and, up to now, only commercially available free-text speech recognition system. DragonDicate is based on what Dragon calls a “Multiple Knowledge Source system design” which integrates a battery of techniques for recognition, including acoustical engineering, a stochastic language model for tackling ambiguities, and a huge active vocabulary. DragonDicate 30k has a base vocabulary of sixteen-thousand words; this can be topped off with an additional user dictionary to thirty-thousand, hence the 30k in the name. To facilitate the entry of new words, the eighty-thousand word Random House Unabridged Dictionary is integrated into the system’s user-interface.

Corpus-based approach
“A major effort has been collecting linguistic data, normalizing it, and obtaining statistics for such things as word frequency and ballistic probability, to enable us to develop huge data tables,” says Baker. “I think we’ve probably analyzed on the order of a hundred million words. We’ve got a fulltime person just searching for large full-text databases for us. And not just in English, either. You don’t know, by the way, of any new cd-roms containing a lot of Dutch text?” she inquires. In addition to the French, German, Italian, Spanish, and Dutch versions of DragonDicate which the company demonstrated at Eurospeech ’91 last fall, Dragon is also working on Japanese and Swedish versions. At the moment, all hands are on deck at Dragon for the German version; the company wants to get it ready for this year’s cebit.
    Requiring a 386-based pc with eight mb of RAM and ten mb of diskspace, and costing just under nine thousand dollars, Dragon-Dicate is no toy. While the company has done its utmost to make it easy to integrate the system into existing configurations — the program uses standard dos keyboard input and can be used with virtually any dos program — DragonDicate appears to have found the warmest reception among the handicapped. Baker tacitly acknowledges this, commenting that whereas an error rate of four percent might be too high for a commercial or industrial application, it is hardly an obstacle for people who cannot use a regular keyboard; indeed, for them, DragonDictate opens a world of opportunities.
    Baker believes that the hardware is finally catching up with the technology, although that in itself could hardly explain Dragon’s success. “When the Intel 80586 chip arrives — do you know how much power that will offer?” she asks rhetorically. “Continuous speech recognition is coming a lot sooner than people think,” she confides. Baker now leans over and says, “now it’s your turn to talk. Tell me, whatever happened to Electric Word magazine? ”