A little heavy-set but not really out of shape, today the Bakers look a lot more like happily aging academics than successful entrepreneurs. But walking through Dragon’s lavish headquarters, it is immediately apparent that they are both. Dragon Systems has grown by nearly 50 percent every year for the past 16; it now employs more than 260 people. Their secret, says Janet, was a decade of self-reliance. Rather than heaping up debt or selling a stake in the company to outsiders, the Bakers insisted that salaries and expenses had to be paid out of revenues. As a result, Dragon focused on solving real-world problems with current technology, and managed to deliver.
The years after Dragon’s hatching brought a laundry list of custom projects, research contracts and first-of-a-kind products relying on the increasingly robust discrete recognition approach. Among the landmarks was Dragon’s first deal, in which a small British firm called Apricot Computers used Dragon’s technology to market the first personal computer to let people open files or run programs by speaking simple commands. (Alas, Apricot had ripened ahead of its time and soon went bust.) In 1986, Xerox workers armed with microphones and radio transmitters used Dragon technology to conduct an audit of the company’s entire inventory of 2.2 million parts.
In 1990, Dragon introduced DragonDictate 30K, the first large vocabulary, speech-to-text system for general purpose dictation. The program enabled a user to control a PC using only voice, and immediately found favor among the disabled, including actor Christopher Reeve.
But Dragon’s discrete technology couldn’t penetrate the general market. Although many people could enter text with DragonDictate faster than they could type, nobody enjoyed being forced to pause between each spoken word. Even worse, competitors were coming on strong with their own discrete speech recognition technology. Everybody knew that what users really wanted was continuous speech recognition, and that the first company to market would be poised to dominate. But everybody also knew that a continuous product was at least five years away, maybe even a decade.
Then sometime during late 1993, the Bakers realized the conventional wisdom was wrong. Knowing the rate at which computer speed and memory were improving, they calculated that top-of-the-line desktop machines should have the power to do continuous recognition within a few years. Just as the pair had once risked their careers on an outlandish new approach to speech recognition, during the first half of 1994 the Bakers started to remake their company in a bid to seize the opportunity and bring their ideas to the marketplace.
While Jim set up a new development team to build Dragon’s first continuous speech recognizer, Janet brokered a deal with California-based hard disk manufacturer Seagate Technologies to buy 25 percent of Dragon’s stock. The company used the cash to staff up its engineering, marketing and sales forces. Within a year, Dragon had the largest speech research team in the world-more than 50 scientists and software engineers.
The new continuous product would really be two programs in one. The first, the recognizer, would go about the actual job of converting spoken utterances into English text. The second program was the interface, connecting the recognizer to both the user and the rest of the computer’s operating system. If the first half was pure science (building on the Bakers’ early work), the second was the frustrating mix of engineering and art needed to turn science into a marketable product.
The trickiest of these real-world issues was making the software run well in a Windows environment. “Windows is awful,” laments Dragon’s Gould, who took on the critical task of designing the user interface. “It’s buggy, poorly documented, inconsistent and pieces of it [are] almost unusable. Yet that’s what all of our customers run.”
By April 1997, Dragon’s team had cleared the key hurdles and started hinting to industry analysts that something big was coming. “We were skeptical,” remembers Peter Ffoulkes of the market research firm Dataquest. Then he saw the demo-which sported a vocabulary of 230,000 words. “We were pretty much blown away with the capability. We didn’t expect it to be here today, and it really is,” says Ffoulkes.
The Bakers had gambled their company and they had bet right. The new continuous recognition product, called Dragon NaturallySpeaking, was an instant hit. Janet Baker’s office began filling up with requests from companies hoping to integrate Dragon’s technology with their software applications. Articles about NaturallySpeaking appeared in publications all over the world; Gould demo-ed the program on CNN. That fall, NaturallySpeaking swept the industry’s COMDEX trade show, winning every major product award.