Maintaining and Increasing Customer Satisfaction Through Voice Automation of Directory Assistance
Several billion times each year consumers call over wireline and wireless networks and pay a transaction fee to get a telephone number. Directory Assitance (DA) Service Providers in North America spend hundreds of millions of dollars to offer this service, with most of the money paying the cost of the operators. These companies stand to save many millions by introducing voice automation. Over the last twelve months most DA service providers engaged the technology suppliers in a process of discovery. By many accounts, 2002 was the year of Requests For Information (RFI) and 2003 is shaping up to be the year of trials and deployments. Telecommunication carriers are cautious in adopting the technology not only because of the revenues that the DA business is generating, but also because it is a point of contact with their customers. Automating DA must produce cost savings without impacting the users perception of quality of service. Their concern is that an automated system may give the appearance of a lower grade of service. By paying close attention to the design of the application, it is possible to introduce automation and at the same time maintain, and even increase the quality of service and along with it customer satisfaction. Emerging Deployments of Directory Assistance In reviewing recent deployments it is worth distinguishing between two modes of deployment: full service and self service. In the full service mode, automation is part of the standard DA offering that is available to callers by dialing a known code such as 411 in the United States. Speech recognition is attempted at the front end of the call, and if successful, the system delivers the requested telephone number. When the automated system is unable to determine which is the requested listing, the call is transferred to an operator. This solution does not entirely eliminate the need for human assistance. However, it reduces the number of operators that are needed to handle a given amount of telephone calls. It does this by completely automating some of the calls, and by offering partial automation (e.g. locality recognition) for most of the non-automated calls. Until recently, this was the only mode of introduction in DA, with a handful of deployments around the world. The great difficulty of applying speech recognition technology to the task of DA kept both the level of automation that was achieved and the number of successful installations low. More recently, weve seen the introduction of inexpensive alternatives to the traditional DA system. Several European carriers, such as Respons in Sweden and Fonecta in Finland have introduced an automated service that is accessible through a separate dialing scheme. In North America, AOL has offered such a service to their AOL by Phone subscribers. These companies offer an inexpensive service by removing the cost of the operator altogether. To the caller, this is the ultimate self-service experience. Telelogue recently deployed a complete DA solution including all 140 million business, government and residential listings with handoff to a live operator serving the local exchange carrier, Blackfoot Communications. The Challenges of DA Automation To understand the challenges of maintaining high quality user experience one has to review the various factors that make voice automation of DA so difficult. First, speech recognition by its nature relies on statistical analysis to determine which one of several alternatives is closest to the utterance that the user produced. The sheer size of the telephone directories makes the task difficult. To try to find a listing in a large locality the system has to match the users request to one of hundreds of thousands of candidate listings. The fact that the telephone listings change daily compounds that difficulty. Second, the listing are often maintained in a format that is optimal for visual scanning but different from the way that users would ask for them; for example the listing could be for Parker, Jeffrey H. MD, but the users might ask for Doctor Parker. Last but not least, more often than not users distort the business name. In fact, from usability studies we find that in 60% of the cases, users phrase their request for a business listing in a way that is different from the way that it is listed in the directory. For example, they ask for Antonios Pizzeria when the business is listed as Antonios Brick Oven Pizza. The Frequently Requested Listings (FRL) Approach To meet these challenges it became a common industry practice to establish a subset of the database that contains only the listing for the business entities that were requested most often. This allowed the technology suppliers to develop smaller and stable Frequently Requested Listings (FRL) grammars, designed to automate only calls made to these listings. The list of FRLs for regular DA is limited to roughly a third of the total number of requests for business and government listings, excpet for Toll-Free DA which has over half of the traffic coming from a few thousand listings. These, in turn, are roughly 80% of the total requests. Thus, the upper theoretical limit for the FRL approach is around 26%. So far, in North America this approach has yielded automation rates that are well below 10%. The Full Directory Approach The Full Directory approach overcomes the grammar size limitation by breaking the task into two steps. The first step uses the underlying speech recognition engine to identify the speech components of an utterance. The second step relies on a middleware that resides between the speech engine and the database and searches the entire database for the presence of these speech components. Using this approach, all of the listings in the directory are candidates for automation, not just those that are requested most frequently. This sets the upper theoretical limit for the full directory approach closer to 100%, and in reality yields automation that is in the 20%-80% range, depending on the constraints of the dialogue with the users. The Voice User Interface for DA Automation Voice User Interface (VUI) specifies a dialogue between the user and a computer. To the degree that the VUI resembles a conversation that one might have with a live operator, the user does not need to learn a new skill. The system does not have to sound like a live operator; it simply has to meet the consumers expectations from a DA call and evoke natural responses. From the callers point of view the introduction of voice automation may be a non-event, as long as the system can deliver the result accurately and reliably. After all, it is easier to deploy a system that does not keep users on hold, has a perfect knowledge of the database and provides a consistent service based on the best methodology. One important element for delivery of customer satisfaction is the ability to accurately recognize those calls to hand-off and pass to an operator for assistance without irritating users with what appears to be too many useless questions. Oftentimes the issue is presented as how many questions do you present to the caller before you hand to an operator? The answer is not a simple number. There is no single level of dialogue that is right for all situations. We view the level of dialogue as a point along a continuum. On one end lies a basic level of automation relying on a small number of questions. At the other end lies a rich dialogue delivering high levels of automation. The choice of dialogue depends on the needs of the customer, and in this case the service provider. Minimal Dialogue Some service providers feel strongly that the best way to introduce voice automation is through a dialogue that minimizes the appearance of call handling by a system. For them, the level of automation that can be achieved through the dialogue is secondary and their choice may be to limit the interaction between the system and the consumer to two questions: System:ACME Directory Assistance, what city please?
User:Red Bank Volvo
System:That number is 732 326 5400 When the system is unable to determine, with a high level of confidence, that there is a one-to-one match between the users request and a unique listing, it would hand off the call to an operator. Based on this foundation the system can grow to automate more of the traffic by introducing additional confirmation and navigation questions. At this dialogue level, many calls that could have been automated are handed off to an operator. If the carrier is open to engage in a more advanced dialogue, more calls will be automated. This is illustrated in Figure 1, and further discussed in the following paragraphs. A Rich Dialogue One of the misconceptions about voice automation of DA is that the answers to the two core questions of locality and listing are sufficient. The reality is that in 20% of the calls, operators have to ask for an additional piece of information, such as the street address. Furthermore, in 40% more calls, the operators are asking for some clarification about the first two questions. On the average, an operator is submitting more than two queries to the database for each listing that they deliver to a consumer. So it is not surprising that getting a user to confirm the name of the business, asking them for the street address, or giving them a choice among multiple listings are all useful for increased automation. The following dialogue illustrates a rich dialogue: System:Blackfoot Directory Assistance, city and state please?
User:Morristown, New Jersey
System:Are you looking for a business, government or residential listing?
System:I have two listings for Dunkin Donuts, one on Lafayette Avenue and the other on Washington street, which one would you like?
System:That number is 732 326 5400 Selecting the Right Level of Dialogue A properly designed voice automated DA system is designed with the level of dialogue as a set of parameters. This way, the number of questions and the user experience can be tailored to the needs of each and every service provider. A major incumbent service provider who is concerned about brand identity is bound to be more conservative. On the other hand, an emerging carrier, who is looking to offer a price-competitive service, is likely to incorporate a dialogue that emphasizes higher automation rate. It is reasonable to expect that even for a given service provider, the level of automation will be increased over time as the consumers are getting comfortable with the automated system. The key is to build a flexible system, and to maintain a dialogue that is easy for the consumer to use. Voice User Interface is the Key for Customer Acceptance This brings us a full circle to question that we started with. How do we maintain and increase customer satisfaction through the automation of DA? The answer lies in the user experience. On the utterance level, the system has to accept the many distortions that people introduce in describing the name of the listing. It must be able to understand and pronounce acronyms such as N double A C P and recognize the peculiar pronunciation of localities (Des Moines Iowa, vs. Des Plaines, Illinois). Finally, it has to accommodate the various natural expressions that people use in response to navigation questions. On the dialogue level, the system has to achieve and project a sense of direction and progress. The study of conversation among humans brought forth the notion of grounding in conversation. What it means is that for a conversation to be successful, the two parties need to keep a common ground and continuously make sure that they understand each other. Similarly, to be successful, a voice user interface needs to emulate this principle and avoid repetitive questions, establish what was understood correctly by the system, and look for additional information that can be used to further narrow the search for the correct listing. Directory Assistance is Ripe for Automation The technical advancements in the field of speech recognition, in general, and voice search technology, in particular, make it feasible for service providers to achieve unprecedented levels of automation in their DA service. Fierce competition in the telecommunication industry, and the need to reduce operating expenses compel the carriers to pursue voice automation. Innovative application design makes it possible to deliver high quality service with speech recognition technology. So when you call for Directory Assistance, be ready to hear a pleasant, recorded prompt asking you: city and state please?
Dr. Amir Mané is the founder and CTO of Telelogue. He can be reached at mailto:firstname.lastname@example.org.