Automation of operator services: a successful application of speech recognition technology

[BThe challenges and design decisions associated with the successful deployment of speech recognition technology in a telephone network-based system providing automated operator services are described. The relevant advances in operator automation that had preceded the present work are briefly reviewed, and the requirements for the introduction of automated speech recognition (ASR) into the automated alternate billing service (AABS) are outlined. The enabling technology advances, including word spotting and barge-in, are highlighted, and the overall deployment strategy for the application is discussed. The billing type and billing acceptance recognition tasks, using speaker-independent whole word models, are described and field results reported. The "bootstrapping" approach to data collection and model improvements is summarized. Next, the current results of the billing detail recognition trials using connected digit ASR are described. These trials are ongoing and involve the recognition of 10-digit telephone numbers and 14-digit calling card numbers. The AABS-ASR service has been fully deployed across the United States and handles over 50 million calls a month for AT&T alone.<>