The Little Dictator

By JOSHUA QUITTNER

My computer didn't understand me. I tried talking to it reasonably, but it was fruitless. When I said, "You say po-tay-to, I say po-tah-to; you say to-may-to, I say to-mah-to," it heard, "Using potato vice, the auto use a tomato." While the idea of potato vice intrigued me, I was getting discouraged by my machine's tin ear. I spent a week with Dragon Naturally-Speaking Mobile ($250), a 4-oz. tape recorder that holds 40 minutes of speech and fits in the palm of my hand. It's designed to take dictation.

I talked into it "naturally," then plugged it into the serial port of my PC, where my voice was magically transcribed into text. The company claims that it can be better than 95% accurate.

People want to talk to their machines. Lawyers and doctors who dictate case notes could use it. So could people who spend a lot of time working outside their office or in their cars. Bruce, the guy who works on my house, uses one to prepare estimates. Brilliant as he is with his hands, he doesn't type.

Another major market is people who type too much and suffer from repetitive-stress injury. According to some estimates, within a few years half of all workers who type on computers will complain of some form of RSI, from numb fingers to inflamed wrists. In fact, I decided to try out Naturally-Speaking Mobile because of Dragon Systems Inc.'s history of making great voice-recognition software, NaturallySpeaking Preferred, for PCs. I've known lots of reporters with burned-out wrists who now dictate their stories and who swear by Dragon's products.

So was I the problem? I talked to Kevin Gervais, who provides tech support for Dragon, and he was mystified. First, he made sure that I had an adequate sound card (check www.naturallyspeaking.com for a list), a minimum of 32 megs of RAM and a 133-MHz or better PC. Since the recorder comes with a good microphone and headset, that wasn't an issue, though it can be if you just use the software on your PC without the recorder. Then we went over the setup procedure. I had read 20 minutes of a Dave Barry book into the recorder to familiarize it with my voice, as required. Next, I had transferred that sound file to my PC, where it crunched away on it for 45 minutes, trying to sort out my way of talking.

An English speaker uses 40 different sounds, called phonemes; the Dragon system used our training session to predict how I would say each of those sounds. Finally, I dumped in a few columns of "vocabulary builder" so the software could learn some of the peculiar phrases I use, such as "trainshouters" and "Bust-A-Move." Everything checked out. "My guess," said Gervais, "is that just by being more conscious of how you speak, you'll improve your accuracy."

He was right: learning to dictate to the machine was trickier than I had supposed. For instance, I pronounce the article "a" as "uh," but the machine understood it only if I said "ay." Also, Gervais admitted, there was a bug in my version (3.01) of the software that cut off the first utterance of any dictation. That bug has been squashed. Now the device is getting better than 90% accuracy and types "tomato" just as I say it.

I'd recommend NaturallySpeaking Mobile to anyone who requires a dictation system. But don't expect to use it right out of the box. It will take more than a few hours to learn how to avoid potato vice.

For more on voice recognition, refer to timedigital.com on the Web. Questions for Quittner? E-mail him at jquit@well.com