The latest jump in PC processing horsepower— Intel's Pentium chip—has spurred the third, and current, wave of voice recognition software, featuring recognition of continuous speech.

No need to pause between every word any more—just speak naturally!

Also, VR prices have plummeted from a few thousand dollars for a kit containing software and a microphone to only a few hundred dollars.

"But it's still not ready for prime time," said MUSC oncologist and technology enthusiast Larry Afrin, M.D. "Not from the perspective of most physicians, at least."

Afrin's demonstration of the current top three packages (from Dragon, IBM, and Microsoft-owned Lernout & Hauspie), is enough to make the technically unsophisticated observer sit wide-eyed and amazed.

Afrin's point was clear, however: the 90-95 percent accuracy claimed by each package still isn't good enough. For most physicians, the time needed to correct errors will still be greater than the time saved by not having to review and edit reports from the transcriptionist.

These systems probably won't let the physician work in another patient, or schedule an earlier tee time.

Using Dragon's NaturallySpeaking system, Afrin dictated a case report at the pace of normal speech (which is slower than how a physician normally dictates, he pointed out), and the computer faithfully displayed each and every word as he spoke it, until.... On the third line, he said "squamous cell carcinoma," which displayed as "screams sell cars in a moment." And the word “Taxol” (a cancer-fighting medication) displayed as “tax all.”

Without touching the keyboard or mouse, Afrin verbally directed the computer to highlight and correct the errant words. The corrections became permanent. That computer would never make that mistake again. To prove it, Afrin dictated, “The IRS likes to tax all people.” You guessed it: “tax all” displayed as “Taxol.”

And the IBM and L & H systems, on the same test, didn't even do as well as the Dragon system.

“The computer can now recognize carefully enunciated phonemes in continuous speech pretty accurately, but understanding context is the big shortcoming in today's VR algorithms (steps in a process),” Afrin said. “Also, most people don't speak in a grammatically correct manner all the time, making it even harder to develop context comprehension algorithms.

“These systems are 90 percent accurate straight out of the box, and if you're diligent about making corrections, they learn and can improve to about 95 percent. Some say 90 to 95 percent is acceptable, but I have to ask, ‘Do you want to stop and correct one out of every 10 to 20 words?'” Afrin said.

He also explained that most physicians talk abnormally fast when dictating, and commonly used phrases are uttered so reflexively and quickly that they sound incoherent to all but other physicians— and experienced transcriptionists.

“In fact, the transcriptionists probably program macro keys for these phrases so that one keystroke types out a whole sentence.” Current continuous-speech systems have a hard time handling abnormally fast dictation.

Ironically, he points out, another problem with VR systems is the immediacy with which the user sees the errors. “When you're dictating the old way,” he said, “your thoughts are focused on the subject matter from beginning to end. But when you're dictating on the computer, you see the mistakes right away, and the natural human reaction is to stop and correct it right then. But that interrupts your train of thought, and it takes more time to get back into the dictating ‘groove.'”

The old way still best?

So why not do it like old-style dictation and wait until the end to look at it on the screen? Afrin explained that because the computer understands only phonemes, not context, words that are misspoken— or misheard—show up as alternative words that have so little relationship to the words around them (“screams sell cars in a moment) that, when you read it later, “you can't for the life of you figure out what you said that resulted in those words showing up.”

Another issue, too, is that doctors are used to dictating into a tape recorder or a dictation system from anywhere, but VR systems restrict them to dictating within six feet of the computer—the length of the microphone wire. Radio microphones might work, but they're much more expensive, and the computer won't receive the sound at the high quality that comes down a wire. Lower quality might mean more recognition errors.

Because of all of these factors, Afrin suggests that VR packages will need to be more accurate than transcriptionists—and just as convenient as current dictation systems—before most physicians will accept them.

“Current systems still make it more the case of the user adapting to the computer rather than the other way around,” Afrin said.

In the meantime, some dictation system vendors are trying a hybrid approach: pre-processing dictation through a VR system and presenting the recognized speech on the transcriptionist's screen as she begins to listen to the dictation—so all she has to do is correct the VR system's errors instead of typing the whole report. Currently such systems are being tested and are expected to be expensive.

Afrin is confident PC-based VR will improve and ultimately save time for the average physician. “It's just going to take more time to refine the algorithms to get a system that handles continuous speaker-independent medical dictation with more than a 99 percent accuracy. In another five to 10 years, the systems might finally be good enough.”

Editor's note: As MUSC rolls out its Electronic Medical Records systems, the prospect of using computers to turn spoken words into electronically stored data to be retrieved in printed form is an attractive one for physicians. For the next few issues The Catalyst will examine this technology, how it's being used and what it promises for the future.