The IDAS project addressed the challenging problem of automating the provision of directory assistance services to the public over the telephone network. Its primary target was to demonstrate the applicability speech technologies1 in particular the applicability of very large vocabulary, speaker independent speech recognition and of speech-based dialog methodologies for cost-effective and user-friendly automated telephone-based directory assistance services. The project officially started on 8. May and was kicked off in Luxembourg on 15. May, 1998.
In the course of the project, two demonstrators have been developed: In the first year, a more system-driven demonstrator was planned and developed that showed the general feasibility of the project. In the second project year, a second demonstrator was developed with better performance and being more oriented towards an actual product. As each country required a different language and its system had to be implemented in a different environment, the demonstrators came as a German, Greek, Spanish and Swiss German system. The Swiss partners were TIK of ETH Zurich, in the role of a technology provider, and Swisscom AG, in the role of a system user.
In the first project year, a number of major achievements have been accomplished. The user requirements for both demonstrators were analyzed and, in order to ensure synergies between the project partners, a common system architecture was agreed upon, allowing partners to re-use parts of their respective previous systems while converging the overall system design. Based on these specifications, the Swiss demonstrator D1.1 was built by TIK. It included TIK's text-to-speech system SVOX which was adapted and integrated into the new architecture, and an open-vocabulary speech recognizer (based on HTK) which allows for dynamically customizable recognition contexts (vocabulary and grammar). Furthermore, an operator module with an appropriate graphical operator console was integrated in order to provide an operator fallback facility. The system was implemented in the Swisscom environment, where it was tested, debugged and improved in close collaboration of ETH and Swisscom.
In February 1999, the Swiss D1.1 system could be demonstrated successfully during the annual EC project review in Luxembourg. In spring 1999, Swisscom, with ETH providing advice and help, conducted field tests to evaluate the demonstrator D1.1.
The second Swiss demonstrator D1.2 was built by ETH in close collaboration with Swisscom, according to a common D1.2 system architecture worked out by the consortium. Based on the experience gained from the D1.1 development, the system architecture specification this time was detailed down to a set of common module APIs, allowing the exchange of modules between the partners and an easy adaptation of the system to new environments and new applications. TIK contributed substantially to these specifications. The usefulness of this concept proved itself for the Swiss D1.2 demonstrator in that the more advanced dialog module of Germany's Temic could be integrated with no major problems into the Swiss system. In the Swiss D1.2 system, some major improvements to D1.1 could be achieved, among others an increase of the vocabulary size from 300 words to 6'000 words, corresponding to a telephone directory coverage of about 1'500'000 entries compared to 120'000 in D1.1, and the integration of a new SVOX version, allowing a much better pronounciation of telephone directory entries containing French, Italian or English person, street or city names. In the field test conducted by Swisscom in spring 2000 in a lab environment and with the operator fallback module turned off, an automation rate of 43% could be achieved.
While an automation of the directory assistance service to a grade high enough to be commercially applicable was not reached within the duration of this project, still significant milestones into this direction were achieved. Integrating the system into a real telecommunication provider's environment would be feasible without major modifications if operator fallback is enabled. Such a system would provide a partially automated service, allowing smooth extension of the automation grade, while offering from the beginning a useful service to the customers.
TIK installed and still maintains an IDAS project web site in order to promote the consortium's work. More information on this project can be found there, at
http://www.tik.ee.ethz.ch/~idas/.
