July 2000

Moore's Law Meets Speech Rec

BY BRIAN STRACHMAN

Also in the 80s, the television show Knight Rider (starring David Hasselhoff) popularized the concept of the talking car. Again, I was sure I would be talking to my car in the near future. Again I was wrong. Prime-time TV viewers tired of the talking car, and David Hasselhoff went on to bigger and better things. Such has been the history of speech recognition -- just when it seems to be right around the corner, it never really becomes mainstream. But now, all that is about to change.

FASTER! FASTER!
Gordon Moore, one of the founders of Intel, is most famous for his theorem known as Moore's Law which states, "The speed of microprocessors will double every 12 to 24 months." (This is usually averaged into the more commonly heard 18 month figure.) His theorem was correct and the past few years have proven to be unique in the PC market.

Until recently, most PC users would see a dramatic improvement in the performance of desktop applications by purchasing a computer with a faster processor. However, processors have become so fast in recent years that the change in response time for most word processors, spreadsheets, and Web browsers has become nominal from one generation of chip to the next. With the possible exception of graphically intensive video games, the PC market is losing its motivation for new upgrades.

This brings us to speech recognition, which is one of the most processor-intensive applications known. For speech rec to work even reasonably well, it takes a very fast processor. The problem in the past was that even the fastest computers didn't do a very good job of dealing with speech recognition. The technology was just ahead of its time, and the few computers that did run a speech rec application were usually prohibitively expensive. However, due to the ongoing advances in microprocessors, the technology is finally reaching a point where speech recognition will reach the mainstream.

NEW TARGETS, NEW APPS
As the need to upgrade computers every few years just to run basic applications dwindles, microprocessor vendors are simultaneously discovering the need to target new markets to sell their wares. As an example, Intel has repeatedly positioned their Pentium III processor as an ideal platform for speech rec and has built in special pneumonics to better handle the technology. Always the supportive partner, Microsoft has also stated that Windows 2000 also has special instructions to make speech rec applications run more smoothly.

While this may seem interesting only to the techies, it is important to remember that these industrial giants have the clout to create new, previously non-existent markets for their products. If they decide speech rec should become mainstream, it will become so. There are three basic platforms for speech recognition: PC-based software, embedded applications, and the telephone.

PC-BASED SOFTWARE
Most people are familiar with the PC products such as Dragon Systems' NaturallySpeaking. This product allows a word processor to write using the spoken word, rather than the keyboard. In the past, products like these were tedious to use, had low reliability, and required over thirty minutes of speaker-specific training. However, with the new technologies the accuracy is extremely high and the training time is around five minutes. Several of Dragon's products actually require a Pentium III due to the special instructions.

EMBEDDED APPLICATIONS
The embedded applications market (composed of cell phone handsets, auto computers and stereos, and home automation) was almost entirely non-existent just a few years ago. Today it holds the potential to generate billions in the near term.

Now that microprocessors are cheap enough to embed in a low-end handset, there is almost no reason not to include speech rec in even the most basic offering. The same holds true for car stereos, home entertainment centers, and various other types of home automation. Very soon consumers will be asking their car radio to switch the channel, and instructing their ovens to "Cook the pot roast at 450 degrees."

TELEPHONE
Finally, there are the products that are most familiar to those in the voice world -- applications that exist remotely and are accessed via the telephone. Already familiar are stock quotes and dial-by-name directories using speech rec. Unfortunately even these high-end products are still sometimes unwieldy. The slightest mispronunciation, slip in grammar, or pause would cause the system to falter.

Nevertheless, the next generation of speech rec applications hold amazing potential. Customers will be able to find the appropriate sales or service department automatically just by asking. Natural language processors will interpret myriad voice commands, questions and complaints, and send the caller to his or her destination quickly and easily.

GLIMPSING THE FUTURE
On the horizon are security products that analyze a voice print instead of requesting an account number and access code. Help desks will increasingly become automated as speech rec systems gain the intelligence to answer common questions without any human intervention. The net result is that call centers will run more efficiently and cheaply. Agents will only be used when absolutely necessary, and then only for more advanced sales or technical assistance. Many of the traditional "galley slaves" of the call center stadium will need to increase their skill set to remain employed.

Not surprisingly, all of these applications still sound like something out of science fiction. These are not new ideas, though. They are old ones simply waiting for the technology to catch up with the minds of speech rec innovators. Some of these products are here now, and some are just around the corner. But the time is right, Moore's Law has been proven correct, and finally the technology is here for speech recognition.

Brian Strachman is industry analyst, Voice and Data Communications, Cahners In-Stat Group. To correspond with the author, please send your comments to [email protected].

[ return to the July 2000 table of contents ]