Unpredictable challenges emerge from the task of moving speech technology applied in laboratories to applications working in the telephone network. This paper summarizes experiences and results achieved from a real-world telephone application that makes use of an advanced user-dialogue and recognition of isolated words using the Continuous Hidden Markov Model approach. The service developed at Jydsk Telefon, Denmark, is named teleDialogue. It has been running as a commercial service for a field trial period of one year (May 91 to May 92), serving more than 1 million subscribers. TeleDialogue enables call establishment to six commonly used telecommunication services in Denmark simply by pronouncing the names of these services. In this way teleDialogue integrates these six services into one that can be reached by calling one specific service number (0020). In order to achieve information on the use of teleDialogue, it is designed to give comprehensive statistics on the use of the service. Furthermore it is possible to monitor and even record user dialogues with the service.
Even if the recognition performance in the laboratory is higher than 95% (SAM standard), the successful transaction rate in the real-world drops to 85%, meaning that 85 of 100 initiated calls are directed to the desired service.
A sample material, obtained by listening to user dialogues, has been selected to reveal the reasons for this performance drop i.e. extraneous speech, background noise, utterance input level and utterances given by children. Each of these subjects are discussed thoroughly in the paper.
Jydsk Telefon is, with teleDialogue, among the first in Europe to integrate speech technology into real-world telephone applications.