We've come a long way since then.

Companies have learned that to get callers to use their interactive voice response (IVR) systems, they have to decide which applications are right for touchtone and which are right for speech recognition.

Consider: A bank's IVR system that uses account numbers to identify customers can retrieve these number more quickly if callers key them in from touchtone phones than if callers say the numbers aloud. But a travel reservations system can recognize the name of a city far more readily than it would take a caller to key in letters to spell, for instance, "Constantinople."

As with any technology that came to prominence in the last decade, the true test of speech recognition is in its application. And speech recognition, for which we'll use the shorthand "speech rec," is effective because it is so versatile.

For example, with an application of speaker-dependent speech rec known as speaker verification, you can potentially eliminate the need for callers to key in or say account numbers. Instead, callers identify themselves with their voices. Most leading speech rec vendors offer speaker verification software. They also offer text-to-speech software, which gives you the option of using automated voices to interact with callers to your IVR system.

In recent years, a whole industry has emerged, ranging from service providers like BeVocal that host speech rec applications to service bureaus like West that offer IVR services. Some companies, like TARGUSinfo, have developed products that use speech rec to automatically verify callers' contact information.

To be worth using in call centers, speech rec has to appeal to the widest range of customers in the right circumstances. That's why this article focuses on speaker-independent speech rec, as it is the most applicable form of speech rec to call centers.

In terms of technology, one key trend is that more speech rec tools are employing statistical models. Back in the 1990s, speech rec applications relied entirely on static collections of words and phrases that needed updating every time callers used terminology or pronunciations the applications were unfamiliar with. Today, leading speech rec vendors offer software that uses statistical methods to figure out what callers most likely are saying, through context clues gleaned from examples of previous conversations.

Another trend is that more speech rec systems use recordings or transcriptions of customers' conversations with customers to automatically generate dialogues for speech rec applications. A by-product of this development is that more vendors provide systems that allow callers to alternate between speech rec and live service from agents.

The trend has significant impact on how call centers apply speech rec. Indeed, the biggest push in call centers, more than determining when speech rec is better than touchtone, is finding the right balance between automation and live service, even within the same calls.

Understanding Usage

The idea that a machine understands what a caller says may be impressive the first time a caller encounters a speech-enabled IVR system. But the more callers have experiences with speech rec, the less speech rec itself becomes the draw for using your IVR system.

"We've moved past the novelty phase," says Bruce Balentine, chief scientist and executive vice president with Enterprise Integration Group (EIG; San Ramon, CA), a consultancy that advises companies on implementing IVR and speech rec.

Automation works best when your reasons to deploy it align with customers' reasons to use it. In the strictest sense, IVR systems don't improve service; they enable customers to conduct transactions without waiting to speak with agents. The goals of offering IVR are to make these transactions more efficient for customers and, in the long term, less costly for your company.

Speech rec furthers these goals. It can coexist alongside touchtone within your existing IVR applications. And for applications where callers look up travel schedules, for instance, speech rec is the only practicable option. Try asking callers to key in the name of a place like Albuquerque from a touchtone phone.

Yet ease of use doesn't guarantee usage. To avoid automation, customers have, or expect to have, a means of escape. Many call centers direct callers to IVR systems before connecting them with agents. The systems, in turn, often place live service as the last item on menus of automated transactions they allow customers to perform. Nevertheless, from their previous experiences with IVR, customers presume they can "zero out," or bypass these systems and speak with agents by pressing zero.

To gauge how frequently and regularly customers use your IVR system, a helpful metric to consult is the containment rate. This rate, which represents the flip side of zeroing out, is the percentage of callers who complete certain transactions through the system.

By examining containment rates during various stages of calls to your IVR system, you can project when customers opt for speech rec even if they have the choice of zeroing out. From an operational and a financial perspective, containment is a key indicator of a speech rec application's success.

As Rob Kassel, senior product manager with ScanSoft (Peabody, MA), observes, "the money you spend on automation actually pays off if people use it."

Transactions and Conversations

Another benefit of analyzing containment rates is that you discover the types of calls, or parts of them, where automation makes sense.

Containment rates even give you insight into routine transactions, like during calls where customers purchase from catalogs. Certain stages of these calls, such as when agents typically gather contact information from customers, or relay information such as pricing back to them, lend themselves perfectly to automation.

Ted Bray, director of product management with Edify (Santa Clara, CA), cites the example of Ticketmaster, which automates the portions of calls where customers provide the dates, venues and names of the events they want to go to before speaking with agents. The information goes to the agents callers subsequently reach.

Yet you wouldn't necessarily want to automate portions of calls when callers place their orders, especially if agents have opportunities to tell customers about additional items that may interest them, or assist callers when items they want are unavailable. Nor would it be a good idea to automate calls when customers ordinarily ask for help with questions or problems concerning your company's products or policies.

You're better off using your IVR system to collect information from callers before giving them the chance to speak with agents. To allow the call to proceed more efficiently, your IVR system can direct this information to agents so that it's on their screens when callers connect with them.

As Pieter Boelhouwer, Empirix's (Waltham, MA) director of voice monitoring solutions, puts it, "If a customer is used to zeroing out, your first goal must be to make the agent more productive."

Bear in mind that during calls where customers choose to speak with agents, service is the most important part of productivity. Loyal customers result in more productive agents.

Whether an agent confirms a sale or resolves a problem, the conversation that ensues can have an effect on a customer's perception of a company. In these circumstances, customers expect agents to relate to them not as automatons that process transactions, but as people. If you want to encourage repeat business and trust among your customers, you have to allow them the choice of speaking with human beings who are accountable for what they say, and how they say it.

By contrast, when customers look up account balances or verify whether checks cleared, for example, they usually care most about speed and convenience. And the more they have to call IVR systems or view Web sites to get the information they need, the more they expect certain types of transactions to be automated.

When retrieving bank account balances, customers generally refer to their banks' IVR systems instead of waiting on hold for agents to provide this information. Containment rates indicate whether the use of an IVR system has become a habit, as is frequently the case for customers looking up account balances.

Containment rates for specific types of IVR applications also reflect how much these applications have in common from one call center to the next. The more frequently customers encounter IVR applications that are alike, the greater the probability that customers develop habits for interacting with them.

Or, to put it another way, the better you can predict what customers are going to say during a particular type of call, the easier it is to automate that call. The same is true in reverse: when customers know what to expect from an IVR application, they're more likely to use it.

Hence another reason containment rates for automated balance inquiries are so high: they often prompt customers to key in similar sequences, like one, one, the caller's account number and the last four digits of the caller's social security number.

As EIG's Balentine says, the hallmarks of an effective application of IVR, with or without speech rec, are "predictability and reliability."

We can now recognize why speech rec is, and continues to be, an efficient tool for automating transactions over the phone, especially travel reservations, stock quotes and movie schedules.

These applications of speech rec are efficient precisely because they're straightforward. They usually ask callers to identify themselves and state what kind of information they want. In return, callers receive the information in a matter of seconds.

Yet just as a tree falling in a forest doesn't make a sound when no one hears it, customers won't acknowledge speech rec systems that fail to respond to them.

If a speech rec system were to greet a caller with "Ready!" and offer no suggestions as to how the caller should answer, you can bet that the caller's next step would be to hit zero.

Callers need to understand the rules that govern their interactions with your speech rec system. The system must give callers a clue as to what words it recognizes, or what categories of information the system is prepared to receive at any given moment.

That leads to one of the toughest questions you'll come across as you implement a speech-enabled IVR system: How much guidance do customers need?