A Change of Interface

Seven of every 10 phone calls customers make to Sears, Roebuck and Co.'s 845 department stores are transferred to sales associates in specific departments.

Sears used to need about 3,000 operators-three or four per store, with only one on duty at a time-to handle the calls. Customers might have waited as many as 20 rings before the operator answered, and one in four calls was misdirected.

To state the obvious, this was not good. And Manager of Store Operations Jim Dutton knew it. So last year Sears, based in Hoffman Estates, Illinois, outsourced its call centre operation, reassigned its operators and launched a speech-recognition system to handle most of the calls going to specific salespeople. Speech recognition, a technology whose time seems to have arrived, now handles 56 percent of 120,000 daily calls with a better than 90-percent accuracy rate, says Jan Drummond, a company spokeswoman.

Sears won't disclose what the project cost. Analysts estimate a price tag, including hardware, call centre software and speech-recognition technology, of at least a half-million dollars. Drummond does say the project paid for itself in a few months. William Meisel, president of TMA Associates, a Tarzana, California-based speech-recognition consulting firm, estimates that Sears saves about USUS$ 96 million a year. He bases that estimate on potential savings from operators' salary and benefits. "Speech recognition is a fundamental change in the phone interface."Companies have adopted speech recognition for internal use among employees or, like Sears, for customers. Department stores, stock brokerages, airlines, auto dealerships, phone companies and package delivery services use speech recognition for various self-service applications. The goal is better service.

There's never been a friendlier user interface, says Bruce Parker, senior vice president and CIO at United Airlines Inc. in Chicago, which uses speech recognition for its employee reservations systems.

Speech recognition can augment or replace existing touch-tone interactive voice response (IVR) systems, whose lengthy menus can be frustrating to navigate.

Some applications are simple-a basic off-the-shelf application directory can store up to 500 names. Applications like the one at Sears are more complicated, involving tens of thousands of words and phrases.

Supporting Self-Service Applications

Speech recognition first appeared in the 1960s when it was developed in university and corporate laboratories. Since the dawn of the personal computer era, using speech recognition to command and control the desktop has been a lofty goal. The ability to dictate documents was another promise. The technology is closer than ever to these goals, but speech recognition has really taken off in self-service applications.

Two things make this possible. The technology has become good enough for the limited vocabularies required by these applications. (Limited is relative: Some bigger applications have vocabularies of 50,000 words.) Second, computing power grew enough to handle the computational demands cheaply.

Details vary, but inside each product is a database of words and phrases that the system can understand when a caller speaks them, triggering certain actions: connecting a customer to a department, giving the status of a shipped package or letting the caller buy a stock. The more complicated the application, the bigger the database of words, phrases and pronunciation variations needed.

E-Trade Group Inc. in Palo Alto, California, which has an online trading Web site, also uses speech recognition to let clients buy or sell stocks over the phone. The application, called TeleMaster, has a vocabulary of more than 55,000 words and phrases. It recognises company names and symbols and their synonyms.

For instance, it understands IBM, International Business Machines and Big Blue.

It also recognises variations and synonyms for everyday words like yes-yup, yep, correct, right, affirmative. Colloquial expressions for yes such as uh-huh and OK and foreign affirmatives such as si and oui will be added as the application expands.

There's no shortage of developers offering products. Sears uses technology from Nuance Communications Inc. in Menlo Park, California. United and E-Trade use technology from SpeechWorks International Inc. in Boston. Nuance and SpeechWorks both offer tool kits and building blocks or objects for developing applications with large vocabularies.

Some developers aim solely at the telephone directory segment. Among them are Phonetic Systems Inc. in Burlington, Massachusetts, and Registry Magic Inc. in Boca Raton, Florida. Phonetic Systems targets large-up to 250,000 names-enterprise directories. Registry Magic sells an automated operator, called Virtual Operator, that handles up to 500 names for US$ 12,000 or on a monthly lease for US$ 700. Other developers include Via World Network LLC, a Minneapolis-based subsidiary of Andersen Consulting LLC that targets the travel industry, and Parlance Corp. in Medford, Massachusetts, which targets telephone directory services.

ROI Measured in Weeks

Although dozens of small and large companies have adopted speech recognition to save money, few applications actually produce revenue. One that might some day is under development at BellSouth Corp. in Atlanta. BellSouth worked with SpeechWorks to develop a Yellow Pages product called VAL (voice-activated link) now being tested in Daytona Beach, Florida. BellSouth customers can call and ask for restaurant listings, news, weather forecasts and sports information.

"We're trying to decide whether an electronic Yellow Pages is a product that consumers will pay for or is simply an access mechanism for our customers," says David Shipps, associate director of business development in the electronic publishing subsidiary of BellSouth. "I suspect it's a little of both. We don't have a business model yet."But most companies are fully capable of justifying the investment in speech recognition by the money it saves. At United, for example, employees, retirees and their family members -- 415,000 in all -- would call the reservations desk to ask for routes and availability for flying standby, which is a company benefit. That came to between 1.5 million and 2 million calls a year, about 5 percent of the total to the reservations desk. CIO Parker says the speech-recognition application now takes most of those calls. The big advantage has come from transferring nonrevenue calls to the automated application, which lets agents serve paying customers more quickly.

The next place Parker is looking to use speech-recognition tools is in the existing AVR (automated voice response) arrival-departure information system.

He does not foresee putting the full reservation system on speech recognition because with all the fare possibilities "it would be too complicated."E-Trade is just one of several brokerages using speech recognition. Another is Boston-based Fidelity Investments. Molly Geaney, vice president of technology in the personal investments group, says Fidelity has two applications based on Nuance's technology. In 1998 Fidelity introduced an application that allows anyone, whether a customer or not, to call Fidelity and get information on more than 3,400 mutual funds by speaking a fund name into the telephone. Fidelity customers can also get account balances and trade funds on the phone.

Fidelity also expanded the application to the company's existing Touch-Tone Express system. Now, instead of having to key in a symbol, Fidelity customers can get quotes and balances and place trades by saying the name of a security or fund. This application takes between 250,000 and 300,000 calls a day.

The company is rolling out a natural language application that recognises and acts on customer orders spoken into the telephone as a sentence ("Buy me 100 shares of XYZ at the market," for instance). Geaney says ROI is hard to figure, but customer satisfaction has been high and now representatives can concentrate on adding value to their services, such as giving portfolio guidance.

United, Fidelity, Sears and E-Trade each spent hundreds of thousands of dollars on speech-recognition applications. Less expensive applications also yield results. JM Lexus, a Margate, Fla., subsidiary of JM Family Enterprises Inc., installed Registry Magic's Virtual Operator. Bill Dowd, the dealership's information systems director, says he was so impressed with a demo that he signed up the same day. Virtual Operator directs customers to the department or individual they request. It has saved JM Lexus at least US$ 40,000 a year, which is what it would have cost to add two operators. The application is about 94-percent accurate, Dowd says, depending on the user's familiarity with it.

"The ROI is in weeks not months," he says. "It was realised the day I didn't have to hire somebody to answer the phones."The Achilles' HeelCompanies using these systems say a big concern was whether customers would want to talk to a machine. Most successful implementations train customers the first time they call in. The early adopters agree that just like ATM machines, it may take time for the public to accept speech recognition, but it's hard to imagine that it won't become common enough to make consumers wonder why companies don't use it.

One factor that makes speech recognition so appealing is the Web. Because most companies have or will build Web self-service applications, the adopters say, they've already got the business databases that can serve double duty for the speech-recognition application. All they need is the speech interface, the vocabulary and a connection to the business database. "We're using some of the same data modules we built for our speech-recognition application for our intranet," says United's Parker. "You have a high degree of code reuse that way."(Bill Roberts, a freelance writer based in Los Altos, California, covers business and technology. He can be reached at