ISCA Archive Eurospeech 2003
ISCA Archive Eurospeech 2003

Improving "how may i help you?" systems using the output of recognition lattices

James Allen, David Attwater, Peter Durston, Mark Farrell

"How may I help you?" systems where a caller to a call centre is routed to one of a set of destinations using machine recognition of spontaneous natural language is a difficult task. Previous BT "How May I Help You" work [1,2] has used top 1 recognition results for classification with much better results when tested on human transcriptions. Classifying using a recognition lattice was found to reduce the gap between results on transcriptions and recognition output. Using features generated from the lattice in addition to the top 1 recognition results gave an improvement in classification of 4% absolute over a baseline system using only the top 1 recognition result. This reduced the gap between classification performance on recognition and transcription by over 25%.