Voice Recognition

Creaking to Critical Mass

By Dan Coolidge

DO YOU remember the scene in one of the old Star Trek movies where Kirk and Scotty have gone back to 20th-century Earth, and Scotty is charged with making some transparent aluminum? Scotty finds a metal shop that has a Macintosh, and goes to use the computer.

He picks up the mouse, and using it as a microphone, tries to speak to the computer to give it instructions. Scotty is bemused that it doesn't respond. He's used to a computer that responds to voice in every situation.

Well, we may not be quite at the ubiquitous use of voice recognition, but we are getting there. Voice recognition software seems to have finally hit critical mass

A few years ago voice dictation software was still in the gee-whiz stage. Demos were impressive, generating "ooohhs and ahhhhs" from audiences -- but as as a real-life, every day production tool, it wasn't quite ready. No matter what version you you tried -- IBM, Dragon Systems Inc., or Lernout & Hauspie -- they all suffered from heavy hardware requirements and questionable accuracy. And then there was the issue of training. Each had Sisyphean training requirements: no matter how much training you gave it, it needed just a little more.

If you got a cold, you had to re-train it because your voice had changed. If you re-arranged your furniture, it seemed the echoes in the room changed enough to throw it off. When the rubber hit the road, it wasn't the tool to to get you to a deadline.

But things have improved. . In fact, I am dictating this article at high speed, having just installed Dragon Naturally Speaking 5 on a new machine and taking five minutes to train it to my voice. I have been using voice dictation for almost a year now in production, drafting patents and dictating letters with my fingers rarely touching a keyboard.

A few figures: I can generally type accurately at about 65 words per minute. By accurately, I mean I can get 65 words per minute, including the time to correct my mistakes (not that I make them, mind you...) With a well-trained voice dictation system, I can roll along at about 120 words per minute. When I'm typing, my hands are tied up on the keyboard.

When I'm dictating, my hands are free to shuffle papers, fidget, write notes and cause other trouble. When I'm typing, I have to go back over my work and proofread. When I'm doing voice dictation, I have to go back over my work and proofread. (It's not magic...)

Good and Bad News

There are some things that voice dictation does very well, and are some things which it does poorly. If you expect to sit back and run your office without ever having to touch a keyboard, you will be disappointed. If you expect to be able to dictate large volumes of text reasonably accurately you should be happy. It does that very well.

Out of the box, most voice dictation software is not prepared for the peculiar vocabulary of most lawyers. Dragon offers a legal specific dictionary, but it isn't yet available in version 5. Although it helps avoid some training, it's not inexpensive (about $400) and has a lot of vocabulary that you may never need. How often do you say "usufruct"?)

A better approach, also offered by Dragon, is to scan a number of your pre-existing documents to learn your particular vocabulary and writing style. It means spending a little bit more time training, but it results in a smaller dictionary and hence a faster operation. The more you train the program, the better it gets at understanding you and accurately transcribing your golden prose.

This provides a natural segue to what has historically been one of the problems of voice dictation: you spend all the time training on one computer and then find yourself at another computer having to start the process all over again. With USB microphones (available from Telex, that plug directly into the USB port and not needing a sound card), you can take your voice training files from computer to computer as long as you bring your USB microphone along.

I have three computers that I use voice dictation on, and three USB microphones. I do most of my training on the one in my office, and bring the voice training files to my other computers and they work fine.

If you are to believe the promotional material, Dragon 5's hype says you can dictate your e-mail, move around menus, answer your phone and do your laundry with voice dictation. Seriously, it puts a little Dragon symbol on screen whenever you are in an application it can support, which does include email. With the top-end version, you can even write macros to automate processes.

I dictate large blocks of text and leave it that. he other functions are easier to do with my mouse and keyboard. And typing e-mail helps keep me terse.

Keep Your Secretary

Voice dictation will not yet replace your secretary. Voice dictation will not relieve you from the chiller out of proofreading. For example, I chose not to correct the preceding sentence. As you can see, it would pass the spell checker but not make any sense. "Chiller out" was meant to be "chore." I have other examples which are more salacious, and which made me a more diligent proofreader.

Of course, training is the key to successful voice dictation. I don't just mean training the program -- you have to invest time to learn how to use the system and how to speak. Clear speaking helps. Pretend you're reading the nightly news -- use your best radio voice.

As importantly, when the program makes a mistake, you need to invest the time to correct it using the correction features of the program. When you first start, you'll find yourself doing this a lot.

If you just type in the correct phrase, the computer will never learn anything. But spending the time in up front training will pay off later when you really need to depend on it. You'll find yourself breezing along at 120 words per minute with only the rare correction required.

If you are somewhat adventurous, most voice dictation systems (at least in the higher grade versions) offer a macro capability. There is something warm and fuzzy about turning into your computer and saying, "Take a letter," and seeing it bring up your word processor, load your letterhead template, insert the date and await your next command.

There is also something remarkably economical about this too. It saves you time, if frees your secretary to do other tasks, and gets the letters done in almost no time.

Would voice dictation doesn't do particularly well is editing. While it has advanced features for moving around a document and selecting text, it's still easier to go back to the keyboard to fix the commas and move words around.

Hardware

As to hardware, there is no such thing as too fast or too much memory. I dictated this on a 733 MHz Pentium III with 256 MB of memory. (Just last week that was a gee-whiz machine.) Attempting to run voice dictation on a slow machine or machine with too little memory is self-defeating.

But think about it: the cost of a proper of computer is about a week's salary for a good legal secretary. In my experience, anything slower than 400 MHz with less than 128 MB of RAM is unsatisfactory.

You'll also want to invest in a noise-cancelling headset microphone. Telex makes a headset USB microphone that retails for about $60. You don't need a sound card with the USB microphone, and as mentioned above, you can do your training once and be done with it. Tabletop microphones (no matter what they say) are too sensitive to noise and changes in the ambient sound.

Even with the best microphone, you'll see your error rate climb if you have a cold, or go from dictating in your office to dictating in the hotel lobby. Voice dictation systems are a lot less sensitive to these changes than they used to be, and recover much faster, but it remains an annoyance.

Go for a headset microphone: tabletop models (no matter what they say) will simply give to problems. Telex makes a headset USB microphone that retails for about $60. You don't need a sound card with the USB microphone, and as mentioned above, you can do your training once and be done with it.

BTW: Though Dragon Dictate was recently acquired by Lernout and Hauspie, the rumor is that in the legal marketplace, the NaturallySpeaking product line will continue, perhaps enhanced by features from the Lernout & Hauspie line.

And so as I said, it ain't Star Trek but it's getting there.

Dan Coolidge is of counsel to Fish & Richardson P.C., in Boston.