Speech recognition

This case-study has kindly been made available by Vocalis Group Plc.

Vocalis develops directory enquiries service with Speech recognition for Telia

Jeremy Peckham

Chief Executive,

Vocalis Group Plc,
Chaston House,
Mill Court,
Great Shelford
Cambridge CB2 5LD


Vocalis has enabled Telia TeleRespons the Operator Services division of Sweden's major telephone company to significantly reduce the running costs of directory enquiries whilst improving customer service. This was achieved through the development of an innovative system to semi-automate the directory enquiries service using a combination of voice response and speech recognition technology. We consider why Telia TeleRespons chose this particular solution, the first of its kind in the world, and see how it has brought about a 8% overall increase in operational efficiency without compromising customer service levels.

The need for change

Telia TeleRespons (hereafter referred to as Telia), is Sweden's leading network services provider and one of the world's most progressive telecommunications companies. Because Telia's prime function is to facilitate and raise the efficiency of telecommunications contacts between customers, its principal commercial assets are human knowledge and modern technology. It supplies a range of operator-assisted services, such as directory enquiries, to its customers and handles over 60 million calls a year. Turnover in 1994 was US$130 million and it has 2,400 employees at 17 locations.

Sweden's telecommunications market is highly sophisticated and callers expect a very high level of customer service. The openness of the Swedish telecommunications market means that established providers such as Telia face stiff competition from new players.

Providing operator services is a highly labour intensive business with personnel accounting for 85% of total costs. Telia recognised that, to retain its leading position, it needed to make savings in this area. Hundreds of telephone operators are required to deal with the steady stream of incoming calls, most of which are very routine enquiries that follow a predictable pattern.

By introducing a semi-automated system it has been possible to reduce the degree of repetitive work done by the operators thus freeing trained staff to focus on more demanding tasks, such as helping clients with complex or unusual enquiries.

The amount of time spent by the operator handling a call has now been reduced by 33%, saving the company approximately 4 million annually.

Why speech technology?

The advent of robust speech recognition technology is the most important development in recent years for operator services. It gives network service providers the scope to reduce costs, improve response to customers, introduce new types of service and improve the working environment of operators.

Commenting on the decision to develop a speech recognition system, Bo Forssten, Vice President of Telia TeleRespons, said, "I believe the main part of what operators do today will be replaced by technology within five years. Speech technology in combination with artificial intelligence and operator fall-back will be the solution."

Telia had to take care to define exactly where in the life cycle of a typical enquiry speech recognition technology could be introduced to ensure minimal disruption to the service offered to the customer while maximising the benefits to the company. As a starting point it studied the main functions of an operator handling a typical directory enquiry call and defined them as follows:

- greet the caller

- listen to the caller's query

- analyse what the caller needs to know

- verify the caller's request

- enter commands to the computer

- release information to the caller.

In order to implement a system that can provide an acceptable service to members of the public, the speech recognition technology must handle certain basic problems in human - machine interaction.

One of the problems is that people will not always wait for a person or a machine to finish speaking before they speak. Another problem is that people will often say more than the machine is programmed to understand.

Vocalis has developed two key features in it's speech recognition technology to overcome these problems. First, word spotting allows key words such as "operator" to be recognised even when they are surrounded by extraneous words, such as "I want the operator please." Second, 'talkover' allows a caller to specify what they want to do while the system is speaking, rather than having to wait for the system to finish. Furthermore, Vocalis 'talkover' ensures that the system is only interrupted by an utterance that contains a valid command, and not by a cough or background noise from the caller.

Using a combination of these features coupled with Vocalis' high accuracy speech recognition, a robust public facing interface has been implemented.

Setting the scene for speech recognition

Critical to the project's success was the development of a speech recognition system that callers would welcome immediately. Working on the assumption that customers do not forget bad experiences, Telia knew that if callers were not comfortable about talking to a computer, they would quickly decide not to use the service at all, which would have disastrous consequences. A major emphasis throughout the project was placed on understanding the human factors aspects that would make the system seem as natural as talking to a person, simple to use and acceptable to the widest possible audience.

From January 1992 Telia embarked on a three-year programme of collaboration with Vocalis, along with local distributor Logica Svenska AB (who provided the systems integration expertise) to design, develop and implement a viable speech recognition system. An initial requirement was to conduct a thorough analysis of all the factors (market conditions, pricing, role of the operators, unions, technology and suppliers) that would determine the success of this ambitious project. From the outset the project had to be meticulously planned and managed.

As part of the market analysis Telia considered a range of variables, for example:

- what do customers think they pay for?

- how do the customers perceive the technology?

- how would customers like the dialogue to be designed?

- are customers prepared to pay as much after automation?

- which services can be improved?

Another major area of consideration was the potential impact on the role of the operators within the company; the unions were closely involved in these discussions.

It was essential to map out a schedule for the introduction of the speech recognition system. As Bo Forssten explains, "Our customers are the ones who decide the speed of introduction of the technology. If we introduce the technology too quickly or if it is of poor quality, we may close the possibilities for the future."

Technical requirements

By the end of 1992 Telia had secured agreement from all the relevant parties to make a commitment to the development of the automated system. With rapid advances in the field of speech recognition it is of paramount importance that an accurate list of technical specifications be drawn up and closely matched to a supplier who can develop and deliver an appropriate integrated solution.

Throughout 1993 Telia, Vocalis and Logica Svenska AB worked to develop a system that could be seamlessly integrated into Telia's existing IT and telecommunications infrastructure. It was based on the Vocalis Network Services Platform, a modular UNIX-based architecture, which allows for expansion to several thousand lines.

Together Telia and Vocalis agreed that the speech recognition technology should meet the following criteria.

1. Accuracy levels of over 95% for the key words to be recognised.

2. Word spotting capabilities so that the system picks up only the relevant words in sentences or in noisy environments - customers do not like to be restricted to saying key words only, and there is often background noise.

3. Talkover capability allowing the caller to interrupt the recorded playout to give another instruction - customers do not like to wait if they know what they want to do next.

4. Capacity to add new words - because Telia offers a dynamic service which is constantly evolving to meet customer needs, it is important that the vocabulary of 'recognisable' words can easily be revised.

5. Upgradeable technology - to ensure that the system can be upgraded as and when new speech recognition features and modules are introduced.

The field trial

By spring 1994 the system was ready to be piloted in a six-month field trial which took place in Sundsvall, a town in northern Sweden. The purpose of the trial was to see how customers and operators would respond to the speech recognition technology and to evaluate its viability in relation to a stringent set of performance criteria. Customers from the towns of Uppsala, Vasteras and parts of Stockholm took part in the test. The operators were split into groups so that it was possible to measure the differences between operators using the new technology and those not using it.

The voice processing technology is used at the beginning and end of calls. The customer is greeted with an individually pre-recorded welcome message from the operator, after which the caller is connected with that operator. There follows a brief period of person-to-person interaction during which the operator searches the database for the number the caller needs, then presses a button to release the number to the caller.

It is during this latter stage, when the number is being played out, that the Vocalis speech recognition technology comes into its own. The customer may issue commands including the words 'repeat' or 'operator' during or at the end of the release. If the customer wants to talk to the operator again, for example, to ask for another number, he is immediately reconnected and the 'new' operator has a log on their display showing what the customer asked for last time. Thus the operator does not need to bother the customer with unnecessary questions.

The results from the field trial completed in November 1994 were very positive. The following results were achieved:

- an 8% increase in efficiency

- 97% accuracy in speech recognition

- the word spotting function worked well

- the talkover function worked well

- the unions recommended wide scale introduction of the technology

- staff were positive in their response to the technology

- 90% of customers were pleased to use the new system

- 6-7% thought the new service was better

Overall these results were a positive affirmation of the system and signalled Telia's approval for it to be scaled up for implementation across the whole of Sweden.

Nationwide roll out

In January and February1995 twelve platforms supporting 360 telephone lines were installed to support Telia's directory assistance services nationwide. The systems offer considerable capacity to be expanded should demand for the service continue to rise. Early signs indicate that the system is running smoothly and that it has been well received by customers.

Telia has met its objectives, maintained its customer base and improved the efficiency of the service. The company is able to deploy staff more strategically and ensure that a human operator is always available to attend to the minority of callers who are not happy to use the system.

A major new feature has been recently introduced to allow the customer to instruct the computer to 'connect' him to the number found. Users will benefit because they no longer need to write the number down and then redial; this is a significant advantage, particularly for the increasing numbers of mobile phone users, especially those operating within a hands-free environment in a car.

The successful deployment of the semi-automated directory enquiries system across Sweden represents a major landmark in development of high accuracy, robust speech recognition technology technology which is able to cope with a wide variety of public users of all ages and accents, as well as different telephone networks.


Return to Top of Page or back to The Call Centre Case Files.