Speech Therapy for the Voice-Recognition Industry

Can the Pentium III cure the stutter of voice-recognition programs?
Publish date:

Could Intel's (INTC) - Get Report new Pentium III microprocessor be the speech therapist that voice recognition needs to stop its market stutters? Could VR be the killer app that actually gives somebody a reason to buy a PC with the P3, despite its high price and infamous serial number feature?

Both Intel and the VR companies are praying for it, the former because the P3's debut has been met by a chorus of loud yawns, mostly because it looks like a lot of horsepower without anywhere to go, at least for most business applications -- except for voice recognition. Consider the Bronx cheers that greeted Intel's latest



Pentium III: Big score ... or big Snore? -- Ziff Davis News Network, Feb. 21, 1999 PIII advances aren't enough -- PCWeek Labs, Mar.1, 1999 Pentium III? Oh, Please! -- Sm@rt Reseller, Feb. 21, 1999

And the VR companies, including

Lernout & Hauspie



Dragon Systems

as well as


(IBM) - Get Report

own ViaVoice program, have everything to win or lose here. The basic problem is that the current versions of their products aren't yet ready for the mass market. Despite voice-to-text accuracy rates that can approach 95%, such accuracy comes with a steep price -- the tedious and laborious process called "enrollment," which is VR-speak for training the computer to recognize the user's voice. Their current software versions call for new users to spend almost half an hour reading standardized text into their PCs. Both Dragon and L&H swear it only takes 18 to 20 minutes, but our hands-on test clocked at nearly 30 minutes in both cases. This is a mind-numbing, dry throat exercise that must be done in a quiet office, free from unexpected interruptions and noise, a luxury for most professionals.

What's more, the accuracy rate after this initial enrollment looks something like this (recorded with Dragon's Naturally Speaking):

"What's more the accuracy rate after this initial role that look something like this."

What about that 95% accuracy rate? It's available only to the extremely diligent -- those disciplined enough to make the effort to continually enroll new words and correct mistakes as they occur.

A Solution in Search of a Problem?

Now along comes the Intel Pentium III with a feature called Single-Instruction, Multiple Data, or SIMD, extensions, which both Dragon and L&H claim will cut initial enrollment time from half an hour to something like five minutes. The SIMD extensions are part of an added instruction set etched into the silicon exclusive to the P3 that provide a sort of parallel processing capability that can off-load multimedia, 3D graphics and some network and voice-processing tasks from the main CPU. For voice processing, which relies on staggeringly immense floating-point mathematical algorithms to pattern and recognize the spoken word, SIMD is the princess' kiss that just might turn these linguistic frogs into the software equivalent of the crown prince for the instant-gratification generation.

"Speech recognition is certainly


good candidate for a killer app that needs the power of the P3 and SIMD," says Keith Diefendorff, senior analyst at

MicroDesign Resources

and editor-in-chief of

Microprocessor Report

. But while the P3 might be the magic touch for voice recognition, Diefendorff says Intel needs a lot more to make the P3 a winner. "In general, right now Intel is desperate for apps that use more processor power," he says, explaining that the current dearth of such programs is due to the need for software companies to reconfigure their applications to take advantage of the new SIMD instructions.

Dragon says the P3 version of Naturally Speaking is done and will be available to the general public by the end of May. L&H is promising delivery of its new VoiceXpress in the first half of 1999.

The Game Boys' Dream Chip

But other software support is sketchy at best. While


(MSFT) - Get Report

says Office 2000 will support the SIMD instruction set, and


(ADBE) - Get Report

says some of its graphics programs such as Photoshop may run better, the only serious rush to embrace it seems to be coming from computer-game software makers.


(ATVI) - Get Report

, for example, has already released optimized versions of

Quake III Arena


Battlezone 2


Heavy Gear 2

, with game developers looking toward video frame rates that are up to 22% faster. The suitability of the P3 for games is more than a little ironic -- given that Intel has positioned it as a business system CPU and the less-game-capable and SIMD-free Celeron as the home machine.

In addition to gamemakers, at least one Web portal is getting in the P3 spirit.



has announced that it is building a 3D Web portal and search engine called Excite Extreme which, thanks to an exclusive deal with Intel, will be available only to P3 users, a fairly small initial market for sure.

More to Come

"Remember that systems with P3s are just now being shipped," says

International Data Corp.'s

semiconductor analyst Mario Morales, who says he expects Intel to aggressively support software developers. By the end of the year, he expects to see $1,500 P3 systems aimed at the mass market as Intel slowly squeezes the Pentium II out of the market.

While Intel hopes to see a thousand points of software lighting up its heavens, reviewers at technical publications as well as corporate IT execs have pretty unanimously agreed that the P3 offers only incremental performance increases for most applications. As Diefendorff puts it: "The P3 is middle of the road: It's evolutionary, not revolutionary, and from that standpoint does not offer the order of magnitude increases required to drive the next level of killer apps."

The Stakes Are High

So for Intel, the P3 looks like a slightly more powerful market placeholder to help counter archrival


(AMD) - Get Report

new K-6 and forthcoming K-7 chips. Of course, if Intel succeeds in making SIMD the de facto standard for 3D and multimedia, it could deal a serious blow to AMD's competing 3DNow technology. This is because software written for 3DNow will not show the same performance increases on a chip using SIMD -- and may not work at all without being extensively rewritten. With Office 2000 optimized for SIMD, this could be the latest Wintel saga that could effectively lock out AMD,


and other chip competitors unless they also embrace SIMD, a task that could require re-engineering and resulting shipment delays.

For their part, both L&H and Dragon desperately need the sort of home run that the P3 promises to deliver. As



Herb Greenberg,

Kevin Petrie and

Jim Seymour have so ably pointed out, L&H has become the corporate poster boy for short-sellers, what with its dicey R&D writeoffs and the resulting

Securities and Exchange Commission


Meanwhile, Dragon has had its own problems with the SEC. Its anticipated January IPO has been delayed by two major revisions (most recently on March 4) of its S-1, adding a slew of new lawsuits -- mostly claiming intellectual property infringements -- as well as stunning new details about how Microsoft's

Whisper program might do to Dragon and L&H what Internet Explorer did to the fortunes of



Perdue helped launch three technology companies in roles ranging from marketing executive to chairman/CEO. He has written widely on technology for InfoWorld, PCWorld, Interactive Week, Forbes ASAP and many others. Perdue is also editor and publisher of

Wine Investment News. At time of publication, he held no positions in the stocks discussed in this column, although holdings can change at any time.