| | | | | |       Listen to Me       | | | | | |

By Hiawatha Bray, Globe Staff, 07/31/97

Like nearly all journalists, I talk too much. I'm constantly talking to colleagues, readers, sources, even to myself. But not to my computer.

So I don't plan on buying the latest speech-recognition products from IBM Corp. and Newton-based Dragon Systems Inc. These systems "hear" spoken words and transcribe what's said into text. Not for me. But if you're a keyboard-phobic computer user, this software could change your life.

Last year I wrote about Kurzweil Voice, a program that did an adequate job of recognizing voices when run on a midrange Pentium computer with loads of memory. But this year's editions have taken a big step forward.

Kurzweil Voice was a "discrete" speech-recognition program, meaning that you had to talk to it one word at a time, with distinct pauses between each word. Talking that way is a drag. What's needed is software that can make out continuous human speech, where the words all run together. That's the way we all talk, and it drives computers daffy.

Or at least it used to. Last year, IBM showed me a continuous-speech product that was designed for use by radiologists. Believe it or not, computers find it easy to understand radiologists because, while they use long, complicated words, they use relatively few of them.

Continuous recognition of everyday speech is harder, because the vocabulary is larger. But IBM has done a decent job of it with ViaVoice, a product that will hit the shelves in late August and sell for a mere $99. I tried a beta version of ViaVoice, and found that it produced tolerably accurate results when used in a quiet room. The accuracy of the product fell off when I tried it in the hurly-burly of the Globe newsroom. So shut the door when you use ViaVoice, or tell everybody else to shut up. In fairness, it must be said that ViaVoice is an introductory product. IBM plans a beefier version later in the year, along with continuing upgrades.

For a more robust and costly alternative, there's Dragon Systems' NaturallySpeaking, which has been on the market for several months now. Dragon has assigned its product a "street price" of $699, but that's far too rough a neighborhood. A quick call to a distributor got me a price quote of $299.

What do you get for an extra $200? One heck of an impressive product. NaturallySpeaking is faster than ViaVoice, and more accurate as well - it made no more "typos" than my fingers do. Indeed, NaturallySpeaking is the first speech-recognition program I've seen that could tempt me to put away the keyboard - at home, anyway. Like ViaVoice, NaturallySpeaking loses some accuracy in an open office setting.

That's not the only problem with these products, either. There's nothing like an hour with a speech-recognition program to remind you of how stupid computers are. If I were reading this column to you, I wouldn't have to spell out every comma or question mark. Humans can figure it out from the context of my words. Computers can't. It's just noise to them.

The programs can automatically capitalize the first word of a sentence, and NaturallySpeaking recognizes proper names. Say "Bill Clinton" and it types Bill Clinton. But to capitalize most words, you must utter an extra clumsy command like "capital-letter," and then the word. It's easier to hit the shift key.

Still, speech-recognition will be a blessing to people like my sister Rosemary, who suffers from a hand disorder caused by years of typing. She plans on buying one of these programs, and she's already purchased a machine to run it on - a husky 200-megahertz Pentium processor with 48 megabytes of RAM. You'll need that kind of firepower, plus a large hard drive, to make either of these programs work. Put the old 486 out to pasture.

Speech recognition won't catch on in newsrooms, or other workplaces where lots of people cluster together. Such places would soon resemble the headquarters of some weird religious cult, with hundreds of workers staring at their computers and muttering a constant neo-Gregorian chant of information. We'd need separate cubicles to block out the noise. But for home users or workers with private offices, the tyranny of the keyboard may be coming to an end.

You can send Hiawatha Bray electronic mail at bray@globe.com.

This story ran in the Boston Globe on 07/31/97.
© Copyright Globe Newspaper Company.