This work was begun when Cox was a Visiting Scientist at Nuance Communications Inc., California in 2000. He worked with Nuance's Head of Speech Recognition, Ben Shahshahani, on developing ways of "routing'' telephone queries automatically to their destination in a company or institution.
The potential benefits of such a technology are obvious to anyone who has used the slow and frustrating systems which are currently universally provided when one telephones a company, institution, government department etc. ("Press 1 for enquiries, 2 for payments, 3 for change of address...") The major challenge in call routing is that the prompt to the customer is deliberately open (e.g. "Please state your query or request", or "Please say which service you would like"). Hence, in contrast to the typical "Please say yes or no" prompts encountered in current voice dialogue systems, the prompt elicits a wide range of responses. These responses can be very different in length, ranging from single words (e.g. "Mortgages") to long responses that may be syntactically and semantically complex or ambiguous, and that may incorporate a large vocabulary (e.g. "There's a transaction on my account that isn't my charge so I need to talk to somebody about getting this removed").
However, the task is made feasible by the fact that the number of possible "destinations'' for a call is usually quite low (<40) and most calls can be unambiguously routed to a single destination. Although high accuracy can be obtained on the majority of calls using information theory principles on keywords, calls that have grammatically complex sentences in them or new words pose a problem. Call-routing is already a successful technology but at the moment, it requires the service providers to have transcriptions of example calls to the system for training purposes, which is a very expensive operation. One aim of this project is to provide automatic routing without transcription of example calls. This work is currently being continued by Qiang Huang (PG) in collaboration with Nuance, who have supplied a database of transcriptions of 13 500 telephone calls. Informal collaboration with AT&T and Lucent Bell Labs in this area is also in progress.
Prof. Stephen Cox , Mr. Qiang Huang