Verizon Voice 411 System Struggling
Saturday, December 25, 2004

Darby Bailey is a 34-year-old California actress. She has three birds and likes to read short stories by Anton Chekhov. And some people in New England do not like her.

It's not that they have ever met her. They haven't and they probably never will. But they talk to her a lot. Sort of.

Bailey's is the cheerful recorded voice you hear when you call Verizon Communication Inc.'s 411 directory assistance on your land line. She is the human end of a new system of speech-recognition software launched by the phone company in September and intended to make the service more efficient. But so far, the system has been struggling.

More often than not, Verizon acknowledges, the system can't find the number it's been asked for and Bailey sorrowfully surrenders the call to a human operator. Although considered pretty skilled for a computer system, she sometimes can't understand what callers are saying despite multiple attempts.

She apologizes a lot. She is relentlessly upbeat. And it drives some people crazy.

"Customer care has definitely decreased for Verizon with Darby," said Meg MacRae, managing partner of The Paisley Group in Castle Rock, Colo., which evaluates the nation's leading directory assistance providers." One big area with Darby is the calls that have to go to the operator because she can't find the number, which means customers are asked to repeat what they want three times. Customers detest being asked more than once, much less twice."

Striking the right tone with customers, not to mention simply understanding them, is no easy task for a computer system. Technicians at Verizon and at TellMe Networks Inc., the California company which provides voice applications for a host of companies, have been continually adjusting Bailey's responses. At first, customers said she spoke too slowly. Bailey, who is described by TellMe as "one of the best voice actors in the world," picked up the pace.

Then, grappling with New Englanders' singular way of talking, Bailey went back into the studio and rerecorded countless business names and locations. And when callers became confused by her apologies for being unable to find a number, her mournful "Sorry" was replaced by, "OK, one moment" before they were whisked to a human.

"Customers thought they were doing something wrong when they heard her say,'Sorry,'" explained Joe Horton, Verizon's executive director of Live Source technology.

Bailey, who used to sing in a band called Eurydice, wound up a voice actress after leafing through scripts her husband brought home from his job as an engineer at TellMe. She admits that the job can be a strain. Recording 40,000 city names has taken months, and Bailey, who sits in a recording studio reading prompts for two hours at a time, admits it can all get a little boring. 

"I try not to think about it being boring, but it can be a test of my patience," said Bailey. "I have to keep my posture up and stay hydrated so I don't slow down. I just try to sound conversational and keep it real so I don't get syncopated. You know, I want to sound empathetic."

And Verizon believes she does; it says many customers think they are talking to a human being. Not that it has been an easy transition. Horton said customer satisfaction dropped by 30 percent in some markets when Bailey was introduced but has risen steadily as improvements have been made and people become more accustomed to the system. Currently, according to Horton, 70 percent of customers surveyed rank their experience in calling 411 as very good or excellent, up from a low of 50 percent.

"One key thing we have learned is that customers really like the automation and they like it as much as the real operator," said Horton. "Obviously, this is a work in progress, but most folks want the number quick and accurate and that is where automation hits. Darby's voice is very engaging and it sounds very pleasant."

Bailey was chosen for the job because her voice "is friendlier than your usual operator, she's the girl next door," said Jonathan Katzman, TellMe's general manager of directory assistance. But Bailey thinks it has to do with her mother. "My mother is a music teacher and she told me to enunciate because I tend to slur," said Bailey. "So I can really enunciate if I need to."

Much of the time, however, the system cannot provide the number. According to Verizon, of the 1.6 million calls seeking government or business listings received each day, the system can take care of 30 percent itself while the other 70 percent are passed along to the human operators. Horton says that "puts us very ahead of the game" within the industry.

Overall, the Paisley group found that the system was doing a good job. Although MacRae declined to compare it's overall performance to other providers, she said that it did better by some measures. For example, it was able to understand the requested location on the first try 88.7 percent of the time.

Jim Smith, a Verizon spokesman, said when the system fails, it is often as much the fault of the caller as the computer system.

"You know, garbage in, garbage out," said Smith. "People don't have good information. They insist it's in Braintree, when it's in Weymouth. They say it's John when it's not."

Either way, some callers tire of the process. Although Verizon said that less than 1 percent of callers attempt to circumvent Bailey by hitting zero as soon as the call gets underway, an unscientific survey by the Globe found some people do just that.

"The system is a complete bollix, as far as I can tell," said Andy Savitz, a Boston business consultant. "For me, it's just a matter of waiting to get to the live person."

Here's how the system works: Bailey takes all callers seeking government or business listings in the 29 states where Verizon provides local service. The requested listing is translated into speech patterns that are matched by the technology to a vast database of possibilities. While the search takes place, the caller hears a curious series of bleeps known as search tones. Verizon spokesman Smith admits they are not actual equipment sounds but "noise intended to give the caller confidence that they are still connected and something good is happening. Dead air would be confusing."

If a match is found, the system triggers Bailey to offer the listing. If not, she says, "Sorry, I didn't get that." In most cases, she tries a second time.

The system was unable to decipher a number of local addresses sought by the Globe. Asked for Worcester, she offered Norton. Asked for the Cambridge restaurant Upstairs on the Square, she first offered Upper Pendleton and then suggested something nearly unintelligible that sounded like, "You said, you've lost your hair?" When asked for the Boston Athenaeum, she handed the call directly to a human.

And that's where The Paisley Group found one of the system's biggest problems. The human operators who pick up where Bailey falls short often aren't listening in, as they are supposed to. The human operator hears a "whisper" of a caller's request and is supposed to be ready with the listing if the caller is bounced to them. But Paisley found that operators had to ask for the listing again in 40 out of 300 test calls, resulting in an average call time of 62.3 seconds, or as MacRae put, "a very long time."

"Customers hate that because they think the operator is not even listening to them," said MacRae. "I won't say the reason why that happens, but there is an opportunity to mitigate customer care issues in an automated environment by working with the humans. I would imagine they do not particularly like Darby."

She's right. Several operators interviewed -- none would use their names -- said that the system created more work because they had to ask the listing again. One, who identified herself as Operator F6145, said that when she used 411 herself, it could rarely understand her.

"I am not a big fan," she said. "I think I am much better."

But Bailey has her supporters, too. Megan Dyer, spokeswoman for TellMe, says that she gets several calls a month from Darby fans. They want to know what she does for fun. (Runs. Works on a screenplay.) What she looks like. (Petite. Blond.) Where she lives. (They're not telling.)