Link to Computerworld

Hands Off: Using Dragon NaturallySpeaking. 

Professional writer finds it twice as fast as keyboarding ...

 By Lamont Wood. (Computerworld Magazine May 18, 2007) 

You decide what you are going to say. You say the words. They appear on the screen. You're done.

That's what writing with speech recognition -- specifically, Dragon NaturallySpeaking Version 9.0 -- amounts to. The lifetime it took to achieve a smooth keyboarding rate of 60 words per minute no longer matters. The skill you effortlessly mastered as a child -- talking -- is all that's required to input text to your computer.

Of course, there's a little more to it than that, but the requirements are trivial compared to, say, a semester of typing class. First, you'll want a fast processor, but everything sold new today will probably suffice. You will want to install all the RAM you can in it -- a gigabyte is a good start. You'll want a fairly quiet place to work, but any place you feel comfortable talking on the phone will probably suffice. And you have to believe that consistent pronunciation is a worthy goal, and not an artificial, elitist imposition.

The software installs from two CDs in a conventional manner. The package also comes with a headphone/microphone, and you'll want to make sure it's connected correctly before beginning -- the correct I/O ports are not always obvious.

After it's installed, the software will examine the vocabulary of your "My Documents" folder, and ask you to read at least one short canned passage while it analyzes the way you pronounce the words. This reading is called the enrollment process and only takes a few minutes -- unlike earlier versions of the software in the last decade, where the process took nearly an hour.

Once it's running, Dragon NaturallySpeaking will install a command bar along the top of screen, with a microphone icon and various menu items. To begin dictation, you position the cursor on the screen, just as if you were about to type something. But instead of typing, you click the microphone icon and began speaking.

At this point, accuracy will probably be about 95%. That means there will be about 10 mistakes per double-spaced page. That's well short of perfection, but the mistakes will be correctly spelled words that just happen to be the wrong words. To correct an incorrect word, you select it and say the correct word. In most cases, that suffices. Correcting the text, in other words, takes hardly more time than it takes to proofread it. Inputting that text, meanwhile, happens at conversational speed, which for most people is between 120 and 150 words per minute. (The software claims to be able to handle 160 words a minute.)

By comparison, the average typing speed for an office worker is about 40 to 50 words a minute. My own is about 60 with an accuracy of 93%, so on the whole, I've found that using speech recognition is about twice as fast as typing. Those who type at hunt-and-peck speeds will experience results that are even more dramatic.

Dragon NaturallySpeaking in action

After the user starts speaking, the yellow "results box" appears near the text insertion point, showing what Dragon thinks it has heard so far. The results will change as the phrase lengthens and Dragon is able to perform further analysis. For instance, it decided that "period" was punctuation in all three sentences.

In the third sentence, it initially placed "to" after "brought" but moments later changed it to "two," successfully analyzing the homonyms. Dragon was less successful with the homonyms in the second sentence, failing to differentiate "which" and "witch."

Dragon types the resulting text after analyzing a phrase, so that what appears on the screen may fall a sentence behind what the speaker is saying (as was the case in this example.) This can be disorienting and users are advised to not watch the screen during dictation.

When finished, the user selected the second "which" and spoke the word "witch" again. Dragon knew enough to respell it as the other homonym.

But when used seriously, the software presents several minor annoyances, the sum of which may drive users back to their familiar keyboard.

First, there's the discouragement of seeing the computer generate mistakes that you didn't make. To counteract that, the user needs to learn how to employ the software's correction functions so that accuracy will gradually improve, until the user's day-to-day vocabulary is mastered. (After six weeks of daily use, my system's accuracy has reached almost 99%.) Basically, when a word is misrecognized, you select it and say "correct (that word)." The software will learn as it goes along, and your user experience will gradually improve.

Frankly, the software does an excellent job recognizing long words, but will stumble over one-syllable words, using an/and, he/me/the or in/on interchangeably. The correction process can tame that tendency, but you also have to learn a different emphasis when proofreading. With typing, by comparison, you have to check the long words but can assume that the one-syllable ones are accurate.

Good at homonyms

On the other hand, Dragon is quite good at differentiating homonyms. "I too took two shoes to the beach," usually came out correctly, albeit after weeks of learning my vocabulary.

Dragon includes a text-to-speech facility to read the text back to you. This will bring out mistakes that your eye would pass over, especially the misused one-syllable words.

Second, you need to stop "thinking with your fingers" and learn to dictate. You have to pretend that you are a BBC announcer and punch out the words clearly and consistently. Slips of the tongue will result in errors, so try not to slip -- or take it personally when Dragon doesn't understand you. Also, you have to learn to pronounce all of the punctuation marks. And while it sounds counterintuitive, not watching the screen while you dictate will make the process go smoother.

Beyond that, to truly master the software, you have to get a feel for its rhythms. For instance, if you want to capitalize a name at the end of the sentence, you need to say it, pause and then say the words "cap that" before dictating the punctuation mark.

On the other hand, keyboards are not going to go away. Speech recognition is a great way for getting your first draft on paper, especially as you can grab the thought in flight and send it directly to the screen, as it were, instead of having to arrest it and then wrestle with the keyboard to get render your thought as text. But editing is easier with a keyboard, especially when you need to move text around. Meanwhile, text does not always read as well as it sounds.

Also, the speed advantage is not automatic, since the speed at which you can mentally compose publishable text is not much faster than typing speed. But when composition is not an issue (as when dictating notes or first drafts) the speed advantage can be dramatic.

Using it in a home office, the chief disadvantage stems from the fact that the cat cannot be convinced you're not talking to it. But that's another story.

Link to Computerworld