Is Voice Recognition Possible?

We've used keyboards to interact with computers for decades now, ever since punch cards went out of style in the 60s. For many people, it's hard to imagine any other means of inputting data into a PC.

Science fiction, on the other hand, shows us a future where keyboards are a quaint memory. Watch any episode of Babylon 5 or Star Trek--the characters don't have Microsoft Natural Keyboards on their desks. Instead, they simply speak to their computers, which recognize their speech and act upon it.

But recent developments in speech recognition mean we don't have to wait for the 23rd century before we can chat with our PCs. Thanks to several new software packages, you can start to put aside your keyboard at last.

Until recently, attempts to develop speech interfaces to PCs had met with little success. Early products have had a presence for years--Newton's Dragon Systems has been around since the early 80s--but none of these products did a passable job, and were mostly used by disabled individuals.

Even when speech programs first started appearing a few years ago, they were mostly limited to "discrete" recognition. In discrete recognition, the computer understands what you say: but. only. if. you. pause. after. each. word.

Voice recognition takes a lot of horsepower on the computer's end, after all, and 386 to 486 class machines just couldn't keep up with the cadence of human speech. But with the proliferation of Pentium technology, home computers have finally become powerful enough to support a more natural rate of speaking.

The latest products to hit the market from IBM, Dragon Systems and Kurzweil all support continuous voice recognition, which means you can speak into the computer without pausing at a normal rate. You can't talk as fast as the guy from the Micro Machines commercial, of course, but a conversational pace is fine.

For the last two weeks, I've been using Dragon's Naturally Speaking on my Pentium-200 with pretty good results. NaturallySpeaking has gotten the best reviews by and large in trade publications, despite its relatively steep price of $199. From my experience, the praise for Dragon's product is well-deserved.

The software has fairly simple hard ware requirements: you need a Pentium-133 or faster with at least 32 megabytes of RAM and an approved sound card. A headset microphone is provided with the bundle.

The first time you run NaturallySpeaking, you go through an enrollment period in which you read a passage from one several books on screen. The computer learns your particular speech patterns by listening to how you're reading the words from the screen.

Once you've gone through enrollment, which typically takes 45 minutes or so, you can run the program. All dictation is done from within the program's mini-word processor window.

To input text into Microsoft Word or Eudora, you have to copy it from NaturallySpeaking and paste it into your other application. While this is an annoyance, it does become second nature fairly quickly.

The bonus of NaturallySpeaking is that the software continues to learn your voice and grow more accurate as you use it. When the computer misidentifies something you've said, you say "Correct that" and type in what you really said. Over time, the system adjusts amazingly to your speech.

If there is one problem with using dictation instead of a keyboard, it's the effect it may have on your writing style. I tried writing a paper for a seminar with the program, only to find that my diction and phraseology when I speak is apparently nothing like my style of writing.

NaturallySpeaking did a fine job of recognizing what I said; it's just that most of us have the system of typing so deeply ingrained in our writing process that it may be difficult to adjust our own cognitive processes.

Despite the adjustments that the program requires you to make in your style, I can't recommend NaturallySpeaking highly enough. It really does change the way to think about your computer.

--Kevin S. Davis '98 is an independent computer consultant and student director of HASC's Advanced Support Team. You can reach him at ksdavis@fas.harvard.edu.