|
Let Your Voice Do the Walking By Lee Hollman
You might recall the old advertisement for the Yellow Pages that advised you to "let your fingers do the walking." Today, that slogan might seem quaint. Speech recognition-enabled interactive voice response (IVR) software enables customers to speak commands, in place of pressing touchtones, to access the information they need. But is the technology as easy to use as it seems?
Consider how frustrating touchtone-based IVR can be. If you've ever navigated a labyrinth of touchtone menu options and still couldn't get the assistance you needed, you don't want to place your customers in the same situation. Speech recognition lets customers ask for what they want. But you don't want customers to endure prolonged hold times if your speech recognition software doesn't understand the requests they make.
So how can you avoid confusing callers when your intent is to help them?
"Our golden rule is know thy customers," says Steve Chambers, chief marketing officer for SpeechWorks (Boston, MA). "Sit in the call center and know their objectives and emotional states."
A little research can help you to design effective touchtone menus that can lead callers to the answers they need. But bear in mind that even the best-laid IVR plans won't please all customers.
"Touchtone remains the most hated form of customer service, but it can be the least expensive form of customer service," says Donna Fluss, a principal at DMG Consulting.
No customer appreciates listening to a series of IVR menus when they want human contact. Fluss recommends that you always let customers abandon IVR menus to speak with agents. But the challenge is to prevent customers from opting out to agents. That's because you save money by allowing customers to get to information or services themselves. Fluss says that IVR menus can be even less expensive than responding to e-mail messages.
Speech recognition software is more expensive than touchtone-only IVR software. But speech rec is also more versatile because it enables callers to make requests or provide information that would be difficult to enter as touchtone data.
If you already have an IVR system and want to upgrade your call center with speech recognition software, Fluss suggests that you don't abandon IVR entirely. "Even though touchtone may be hated, people become comfortable with it," she says. "Don't take away from customers what they're accustomed to using. But please feel free to enhance it."
Fluss explains that the recent recession forced call centers to cut costs by reducing their labor forces. But those centers will need to respond to rising call volumes during an economic recovery. Speech rec lets those centers automatically handle more calls, without hiring additional agents, than they could by using IVR alone.
Below, vendors review their speech rec and IVR products and describe how to best use them. Not surprisingly, they had a lot to say.
IVR That's Up To Par
Although you can depend on IVR systems to provide customers with information, you don't want to provide too much at once.
"There's a tendency [among companies] to have an IVR [system] say something they'd never have an agent say," says Ron Owens, director of product marketing for Intervoice (formerly InterVoice-Brite; Dallas, TX).
He cites as examples IVR systems that run greetings for 45 seconds or that burden callers with lists of instructions and legal disclaimers before they can get to the first menu.
To avoid common IVR pitfalls, vendors recommend that you seek outside assistance. Steve Feldman, CRM specialist for Avaya (Basking Ridge, NJ), says he's seen varying results whenever customers create their own scripts.
"This is not an area to cut corners on," warns Feldman. "Whether it's a consultant or a company that designs the [IVR] scripts." That's why Avaya, like many other vendors, offers their own consultation services to help you design IVR scripts.
IVR scripts comprise the content of each touchtone menu. To write effective scripts, you must give callers the broadest possible choices with the first menu and offer more specific options with subsequent menus. Callers can then quickly narrow the focus of their search.
"Touchtone application design best practices [that call for] no more than five options per menu and no more than three or four levels of menus for any function have proven over time to deliver optimal performance," says Mark Skoog, director of marketing for Syntellect (Phoenix, AZ).
Skoog suggest working with a small, manageable focus group of approximately ten people to efficiently test IVR scripts.
"The idea is not to spend a long period of time trying to analyze statistics and put together a big plan," he says. "The idea is to test [your menus] quickly and modify [them] from there."
You might also hire an outside consultant to organize a focus group that can provide an unbiased opinion on your IVR script's user-friendliness. Otherwise, Skoog suggests looking inside your organization and outside of your call center.
"Don't find people who are already familiar with your scripts," he advises. "They have preconceived notions about how the scripts should work."
Most customers can probably use touchtone menus to handle basic tasks, like finding the price of a product, without opting out to receive live help. But reducing the number of calls that require agent assistance doesn't mean that you need to downsize your agent staff.
"A successful IVR script is not one that reduces head count at your call center, but one that increases the business levels of the organization," says Fred Manus, marketing manager, Nortel Networks (Brampton, Ontario, Canada).
To illustrate his point, Manus recalls how a major brokerage used an IVR system with a speech recognition engine to provide stock quotes to callers. He says that because callers find stock prices more quickly by requesting them from an IVR system than from agents, callers began checking more prices than they had before and did more trading with the firm. But the firm kept every agent employed at its call center to provide customers with additional assistance.
Start Your Speech Recognition Engines
Speech recognition engines are as essential to a speech-enabled IVR system as a mechanical engine is to your car. You use the speech rec engine to build apps to handle the different ways that customers phrase requests. But you don't need to do this app development alone.
"It's rare that you'll find a call center manager and his development team creating [speech recognition] apps from the ground up," says Marcello Typrin, group product marketing manager for Nuance Communications (Menlo Park, CA). "So it's best to work with a vendor directly."
When you're browsing for a speech recognition engine, consider asking vendors about what consulting services they offer. And before you license a speech rec engine, confirm that the vendor keeps its technology current.
"You'll want to find out how often vendors upgrade their engines," says Tom Falk, director of marketing, Philips Speech Processing (Dallas, TX). "You don't want to be developing apps for an engine that the company's moving away from."
SpeechWorks' Chambers advises that you also learn how well speech recognition engines integrate with touchtone IVR systems.
"Two separate engines will vary significantly on different IVR platforms," he warns, adding that the platform can affect the speed and accuracy of an engine. So ask vendors for information on how well their engines work with specific IVR systems. These are usually available in case studies or white papers from speech rec vendors.
A Is For App
When you design speech-enabled IVR menus, you need to write apps that enable callers to select options to hear specific information. And you need to create apps that let you automate tasks, like routing customers to appropriate menus or agents. How well your apps work determine how effectively you deploy speech-enabled IVR at your call center.
How can you design or customize apps to help customers as efficiently as possible? Charlie Rabie, vice president, customer self-service products, Aspect Communications (San Jose, CA), emphasizes "human engineering" over the technical kind (i.e., don't rely solely on your IT staff). Consider consulting product experts - or agents. No one is in more frequent touch with customers than the reps who speak with them daily. Organizing apps can be a collaborative effort.
That's not to dismiss the technical side of designing apps. Although many vendors include proprietary programming languages with their software for writing touchtone- and speech-enabled IVR code, a new standard for apps has emerged.
"What we're seeing is a transition," says Nuance's Johnson. "VXML is an evolving standard. The goal and the theory [behind it] is to remove vendor dependence."
VXML (Voice eXtension Markup Language) is similar to HTML in that it's a Web-based markup language comprising tags that you can add to a text document. VXML, which integrates with speech recognition engines, supports telephone access to Web services. It also supports Web browsing and Web site interaction using touchtone and speech recognition.
Designing apps with VXML permits you more freedom to customize apps than you would have with a vendor's proprietary language. But VXML-based app development can also be time consuming.
"If you have to write [code] from scratch, it'll take you forever," warns Avaya's Feldman. "You'll sit there typing line for line, every syntax of the code." That's why the majority of recent IVR and speech recognition software support VXML in addition to their own languages.
But VXML isn't your only alternative to building apps with vendor programming languages. SALT (Speech Application Language Tags) is also gaining momentum among call centers. SALT comprises tags that enable you to add speech recognition and features to HTML code.
SALT lets you integrate your company's Web site with IVR systems and speech recognition to enable customers to hear text from your Web pages through the phone. If a customer asks to learn more about a product, that customer can listen to a product description posted on your company's Web site.
How do you decide whether to develop apps with VXML or SALT? First, consider the reason that both languages were created.
"VXML comes from a desire to standardize the IVR world," says Heiko Rahmel, vice president of professional services, Philips Speech Processing. "SALT comes from the desire to let Web developers add speech into their Web pages."
So if your goal is to create improved touchtone menus, VXML might be the better bet. If you want to include on-line content as part of your customer service, then SALT is for you.
Since more call centers use IVR systems than speech recognition software, VXML has found rapid acceptance among them. But expect SALT to become equally popular as more call centers adopt speech recognition technology.
"I believe that SALT will gain momentum," says Feldman. "But it's not widely deployed in thousands and thousands of locations. So I would not yet declare a victory."
And remember: Sometimes the simplest solutions work best. You may not need to write apps with VXML or SALT, depending on your needs. To create basic touchtone menus with a limited number of options, the application development modules that come with most IVR software should suit your purposes.
Let the Software Speak For Itself
Aspect Communications provides two options for developing IVR apps with Aspect Customer Self-Service (CSS) 7.0, due for release by January 2003. You can write IVR scripts in VXML code or build scripts by assembling icons into a tree diagram. The icons represent options that customers can select from each touchtone menu. Customers can also perform specific functions, like searching a database for information.
Also new to 7.0 is its ability to integrate with speech recognition engines without voice boards. Rabie explains that servers running speech recognition engines contain voice boards that relay spoken commands from the engines to IVR apps. Since CSS 7.0 communicates with the speech recognition engine directly, you can set rules for routing customers without writing VXML or Java apps.
Looking to give your IVR apps a test drive before your customers use them? With Avaya's Avaya Interactive Voice Response (IVR) 9.0, you can record IVR scripts as .wav files from a PC microphone. After you finish recording, you play back scripts and edit or re-record them as necessary.
You use Avaya Interactive Voice Response Designer, a component of Avaya IVR, to write IVR apps in VXML or with Avaya's proprietary scripting language. You create diagrams with icons that represent each option that customers select from touchtone menus. You can then test completed apps by entering touchtone data from your PC to run through every menu you create.
Omvia from Intervoice also provides an application development environment, InVision Studio. InVision Studio lets you author IVR scripts with Omvia's VXML or with Intervoice's own proprietary language. Depending on the language, you can work from a VXML browser or from a browser for Intervoice's language. The J2EE-based module also lets you develop IVR scripts in Java, using your code or code written with Java-based tools like Dreamweaver.
Intervoice's Owens recommends that users have some programming background. He explains that the software contains libraries with samples of VXML and proprietary code that programmers can piece together to assemble or modify an app.
Unique to Omvia is an optional module that generates reports containing information like the number of callers that do and do not complete a successful search of IVR menus for the information they need; and the reasons why their searches fail.
After acquiring Periphonics in November 1999, Nortel Networks began offering the Periphonics line of IVR products under its own brand. Media Processing Server (MPS) 100, Media Processing Server (MPS) 1000 and Voice Procession Series Information Server (VPS/is) each comprise software for designing and running IVR apps that come installed on Windows NT or Unix servers.
Media Processing Server 100 is targeted to small- and mid-size call centers. The system enables 12 callers to use touchtone menus simultaneously, but you can scale upwards to accommodate 60 callers. VPS/is can handle from between 48 to 240 callers. Media Processing Server (MPS) 1000, intended for large enterprises, provides simultaneous touchtone menu access to 10,000 callers.
Nuance introduced enhancements to its engine with the release of Nuance 8.0 speech recognition engine. The software's new Listen & Learn suite of features includes Auto-Tuning, Dynamic Language Detection and Personalization Kit.
Auto-Tuning helps you increase Nuance 8.0's speech recognition accuracy by enabling you to set the engine to recognize regional or international dialects and accents. You can also set thresholds for background noise so that the engine doesn't mistake the noise for human speech.
The Personalization Kit uses acoustic models based on millions of actual calls to identify the genders of callers, the phones they use (standard or wireless) and the level of background noises. Based on this information, Nuance 8.0 provides callers with customized speech-based interactions.
Rounding out the Listen & Learn features, Dynamic Language Detection automatically recognizes which language callers speak and plays IVR menus to them in that language. Dynamic Language Detection also lets callers switch between languages at any time during the call.
Also new to Nuance 8.0 is Say Anything. If customers use imperfect grammar, Say Anything recognizes key words and phrases to direct them to the appropriate call path. So if a customer says, "I'd, uh, like to, um... buy a Sony TV, okay?" the software interprets the sentence to mean, "I want to purchase a Sony television."
To increase the speed and accuracy of 8.0's speech rec capabilities, Nuance introduced AccuBurst. Since speech recognition engines don't use all of the memory and CPU power of the servers during lulls in call volume, AccuBurst taps unused resources to analyze more thoroughly what customers say and to reduce the time required to do so.
SpeechPearl 8 from Philips Speech Processing also includes improvements. The new engine hastens speech recognition times by comparing what callers say to every word you save to the engine's vocabulary. Previous versions of SpeechPearl could only check a spoken utterance against one million words at a time; the current version can check against several million words at once.
SpeechPearl 8 also pinpoints and ignores background noise that interferes with customers' calls. A new component of the engine lets you fine-tune how much of the background noise it picks up so that you can minimize the chance of its mistaking noise for words.
SpeechWorks automatically allocates server memory for improved speech recognition capability with OpenSpeech Recognizer. The engine's Finite State Transducer (FST) feature lets you efficiently store complex grammatical structures in memory so the engine can as easily recognize a sentence ("Please retrieve today's stock prices for Microsoft") as it could a phrase ("check Microsoft stock prices").
To speed developing speech recognition apps for specific customer requests, DialogModules enable OpenSpeech Recognizer to recognize words and phrases that customers say frequently, like zip codes or product names.
OpenSpeech Recognizer's new Learn feature enables the engine to recognize different pronunciations of the same words. Learn can also generate new pronunciations of words that customers use. To help OpenSpeech Recognizer accurately analyze each call, the Endpointer feature determines when customers speak.
The Vista Interactive Media Response (IMR) IVR system from Syntellect lets callers select how they want to receive information that they request. A caller wanting to return merchandise can listen to instructions through the phone, receive an e-mail containing the instructions, or both.
The current version of Vista Interactive Media Response lets you design IVR scripts in Java and VXML. To enable Vista IMR to integrate with a Nuance or Philips Speech Processing engine, Syntellect also offers Vista Advanced Speech Recognition software.
Unveil Conversation Manager from Unveil (Waltham, MA) is unique in that it's an IVR system with a natural language understanding component. The Meaning Recognition Engine (MRE) works with speech recognition engines to increase recognition accuracy.
Peter Durlach, president of Unveil, says that speech recognition engines can recognize phrases and sentences based on the statistical likelihood of callers using certain words together. But what if a caller says something unexpected? "When the engine hears 'Dow Jones industrial,' it's pretty confident the next word will be 'average,' even though it could be 'rhinoceros' or anything else," he says. An unexpected turn of phrase can complicate how efficiently the engine identifies, and responds to, what the caller's trying to say.
After speech recognition engines interpret what callers say, Unveil's Meaning Recognition Engine analyzes the meaning of each word. If a caller says, "Dow Jones industrial rhinoceros," MRE examines words one at a time to conclude that the caller wants information from Dow Jones & Company. Based on that information, Unveil Conversation Manager can answer with the correct information or route callers to a touchtone menu offering options to find information available from Dow Jones.
Unveil Conversation Manager also automatically reviews transcripts of customer calls and recognizes the grammatical patterns that callers use when making requests. Unveil Conversation Manager then codes these patterns in VXML to recognize these patterns when customers use them again. If you offer product return services through the phone and have several callers asking about "returns," Unveil Conversation Manager recognizes that "returns" (like "exchanges" and "bring back") mean that callers want to return a product.
Who to Speak with About IVR and Speech Recognition
As useful as touchtone menus can be, you sometimes need human contact. Here's how to contact the vendors included in this feature.
Aspect Communications
Avaya
Intervoice
Nortel Networks
Nuance
Philips Speech Processing
SpeechWorks
Syntellect
Unveil
|