Language Barrier

Voice-activated dialing is becoming as much a staple as voice mail. Good thing today's versions are a lot more robust and user-friendly.

The 2000 GSM World Congress was filled with people muttering into their lapels. Why? Hands-free headsets, with microphones built into the ear-bud cables at mouth level, were all the rage. The hardware dovetails nicely with the trend toward voice-activated dialing (VAD), which soon should be as common as caller ID. Samsung, for example, will include VAD as a standard feature in all its handsets, and other vendors are launching handset- and network-based solutions that promise to end yesterday's fussy interfaces.

The timing couldn't be better. In a 1999 Yankee Group survey, 45% of wireless subscribers showed interest in VAD, which beat out staples such as voice mail. VAD also provides access for disabled subscribers, and high VAD penetration industrywide could be a pre-emptive strike against bans on the use of phones while driving.

Another reason why VAD has to work well is that it's an entry point for other, more lucrative voice-driven services such as wireless-data portals. Some VAD solutions already allow providers to program commands that take users directly to, for example, customer service or prepaid balances. Add in location technology, and the potential applications mushroom.

"We're beginning to see carriers explore models around sponsorships to have placement on that pre-programmed list," said Brian Kaplan, telecom product manager at Nuance Communications, a VAD vendor. "Say a major pizza company wants to have its brand be on that speed-dial list. The user would say, 'pizza,' and be connected to that chain's local shop."

User- & Provider-Friendly
VAD is aimed partly at people on the go, so chances are they'll be using it in some fairly noisy environments. Two challenges are distinguishing the speaker's command from background noise and knowing when he's making a request and not just chattering away. The user interface can help overcome both.

"It asks, 'Can I do something for you now?' Then there's a tone, and you speak," said Dan Johnson, Lucent Technologies product-marketing manager. "Rather than keeping the recognition resources listening to you constantly where some background noise might throw it off, we don't start listening until that beep."

Today's network-based VAD solutions also are easier on providers. In the past, most solutions had each call hairpinned through the VAD platform.

"It was very inefficient from a trunking standpoint," Johnson said. "We're using an IS-41 feature called redirection directive, which lets us get out of that hairpin. You dial a star code to access the service, and you're connected to the voice-dialing platform. You say, 'Call ______,' and it says, 'Calling ______,' connects you, and the VAD platform drops out. It means that you can put more users on that platform probably by a factor of at least eight, and it drastically reduces the trunking costs."

One advantage of a network-based solution is more diagnostic tools.

"There's the opportunity to collect data on an ongoing basis to monitor the system performance and then use that data to improve performance," said Nuance's Kaplan. "With handset-based, the carrier has no ability to know whether users are using it with success or frustration."

With handset-based solutions, the VAD technology resides entirely in the phone. Some commands, such as "dial" or "call," might be preprogrammed, but the user has to "train" the phone: He enters a person's phone number and then says the name, which the phone records. Later, when he says, "Dial Jane Smith," the phone looks for that "sound" in its library and, if a close match is found, dials the number.

Because the handset has a limited amount of processing power and memory, VAD can have difficulty discerning between two phonetically similar names or distinguishing the request from background noise.

"It's not that algorithms aren't good," said Muzibul Kahn, Samsung Telecommunications senior director, product management & engineering. "You need more processing. The key is how can you utilize the processing power available in the phone."

Electronic Polyglots
A VAD solution has to be flexible enough to accommodate a wide variety of speakers, and that requirement isn't limited to melting pots such as New York City and Toronto, where it's common to hear dozens of languages, dialects and accents on any sidewalk. Sprint PCS, for example, is adding two Spanish-ready VAD handsets nationwide.

Developing a VAD solution just for, say, English, involves accommodating multiple accents even for a simple command such as "dial."

"If I'm a native Texan, I'll probably speak it a different way than a Bostonian," Kahn said. "Given all those variations, the phone still will be able to extract what it's looking for. It's a much more sophisticated algorithm."

Add more memory and processing power, and the chances of understanding the speaker's request improve significantly.

"We've had engineers look at transcribed data from deployments," said Kathy Frostad, Nuance Communications director of telecom marketing. "In listening to them, they can't understand what somebody's saying, but the recognizer is easily able to accommodate that."

Quality in any Language
Providers that offer VAD say that subscribers seem to understand the technology's limitations and will make some adjustments, such as rolling up the car window when asking the phone to dial. That's helpful because user error can make or break VAD.

"What you find is that people do the initial set up in an office or home, where the background noise is minimal," said Michael Coad, Sprint PCS vice president, subscriber equipment, product realization. "Then they'll attempt to use it in a car or outside, where there's a lot more background noise. That's where you'll find a bigger problem with recognition. Because that's pretty common, we're working with handset manufacturers to make the algorithm a lot more robust. If you were to try it in the latest phones, you'll notice that the accuracy is a lot better in that scenario."

Handset-based solutions typically limit the list to about 25 names. Although that's mainly because of the limited memory available, there is a side benefit.

"After that, you tend to forget who you have in there, whether you put in their first name, last name, 'Mike on mobile phone' or 'Mike at home,'" Coad said. "After you've reached some number, it becomes a little too complicated to continue adding more names and numbers."

If a subscriber complains that the handset "forgets" names, one answer is to suggest that he review his list to verify how they were originally entered.

"In the implementations that we have, you can review all of your programmed name tags," Coad said. "Make sure that you're saying the right name."

Providers can improve VAD performance by improving overall quality of service. For example, if coverage is poor, VAD performance can be poor.

"Speech recognition's ability to work well is very dependent on the signal quality," said Nuance's Kaplan.

Another option is to offer hands-free kits, even if they do attract puzzled looks.

"Some headsets have mikes that are much more directional and therefore cut out a lot of the background noise," Coad said. Boom mikes "are very directional, so they cut down a lot of the noise."