"New Technological Developments - PCs and the Internet"
ARE YOU TALKING TO ME?
SEVERAL NEW SOFTWARE SYSTEMS PROMISE TO LET YOU CONVERSE WITH YOUR COMPUTER. BUT HOW WELL WILL IT LISTEN?
BY BRAD STONE
CAPTAIN KIRK DID it. So did "2001'' astronaut David Bowman, the "Lost in Space'' Robinsons and the cartoon Jetsons. They spoke to their computers and robots, and the machines actually understood and spoke back. After all, it was science fiction.
But suddenly it doesn't seem so fanciful. The burgeoning technology of speech recognition promises to arm our deaf-mute PCs with the ability to act on commands spoken into a microphone. "Speech will eventually replace the need to use the mouse and keyboard,'' says Michael Dertouzas, the director of MIT's computer-science lab. That kind of enthusiasm has spawned sanguine forecasts of $1 billion in overall speech-technology sales by the end of the century and fueled a rush to get software onto store shelves. Do the products now available live up to the hype? Not quite yet. But eventually, you may rely on speech-recognition software to surf the Web, check e-mail from your car--and perhaps even do your banking.
Today the most heralded application of speech technology is so-called "continuous speech dictation,'' in which a computer transcribes everyday dialogue. (First-generation dictation programs, which debuted in 1990, forced users to pause awkwardly between words.) Since individual accents and diction vary, new users must spend about 30 minutes reading to the computer from a training routine. The software's accuracy gradually improves as it is used, adding unrecognized words to its vocabulary and learning from mistakes when they are corrected. It typically takes a few trying days of regular use to achieve the 95 percent accuracy touted on software boxes.
The two packages now available are from IBM and Dragon Systems, based in Newton, Mass. Dragon was first in stores last summer, and its software is still the strongest overall. Dragon Naturally Speaking Preferred ($150) has the best editing features, letting you verbally select botched words and then speak or spell out the correction. It can also read the text of a document or an e-mail out loud, using a synthetic voice that sounds like an intoxicated British butler. Dragon cofounder Janet Baker says her software can transcribe up to 160 words per minute. NEWSWEEK maxed out at about 100--still better than the average of 60 wpm that can be typed on a QWERTY keyboard.
Spoken commands: IBM's ViaVoice Gold ($149) beats Dragon in at least one respect: it lets you skip among the programs on your desktop with spoken commands (example: "open solitaire''). But once you get into your word processor and start transcribing, the ViaVoice editing features are frustrating. You have to use the mouse to click on mistakes and then retype the words, which is maddening when the software is still new and mauling every other sentence...
So whom are these programs suited for? They're not for wimpy computers: you need at least a 133-MHz Pentium processor and 32 megabytes of memory. And there's no reason to endure the initial frustration if you're comfortable with the keyboard and mouse. Also, users who frequently change the subjects they write about will have a hard time getting the dictation software to keep up. "As soon as it gets zeroed in on one thing, you are changing the topic and suddenly throwing words at it that it won't understand,'' says Bill Meisel, the editor of Speech Recognition Update newsletter. On the other hand, it's perfect for professionals like doctors and lawyers who stick to one lingo, and all three companies offer software with specialized vocabulary. Users suffering from a disability, repetitive stress injury or just a serious hunt-and-peck typing deficiency should find the software a major breakthrough...
But to really fulfill its destiny and revolutionize computing, speech needs to move far beyond the desktop. Since 1994 a service by Wildfire Communications has let customers verbally instruct a friendly-sounding electronic telephone assistant to place calls or read messages and e-mail out loud...
The newest systems, now in trial, attempt only to understand the key words in a customer's request, rather than trying to recognize every word spoken. For example, the Texas-based USAA Bank is now testing a natural-language call router by Lucent Technologies intended to replace those awful push-button systems ("push one for customer service . . .'') Customers who call the bank are greeted with the phrase "How may I direct your call?'' They can then make a randomly phrased request, like "I want to open a checking account.'' The computer looks for the words it understands, analyzes their place in the sentence and then routes the caller to the appropriate department--or transfers the request to a live operator if it's confused...
(© Newsweek; U.S. Edition - March 02, 1998 (shortened))