Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Automated Query Identification in English Dialogue

Mark Terry (1), Randall Sparks (2), Patrick Obenchain (3)

(1) Audiologic, Boulder, CO, USA
(2) US WEST Technologies, Boulder, CO, USA
(3) Dept Linguistics, University of Colorado, Boulder, CO, USA

A question-acknowledgment statement classifier (QUASI) was built for use in an automated 'street-map directions' dialogue system. The system uses a cepstral pitch extractor to construct a pitch contour, which is then processed by a set of shape detectors. This information is sent to a rule component to classify the statement type. Using cellular telephone voice samples, basic query and statement types, where rising or falling pitch is the predominant cue, were classified with an accuracy of 90%. A Neural-Net classifier, trained with a back-propagation algorithm, gave a recognition accuracy of 85 % using the same pitch contours. Statements such as, 'okay' and 'uh-huh', which are employed to acknowledge correct receipt of an instruction and to signal readiness for further information, are commonly produced with a intonation rise. We tried to identify this statement type by utilizing a pitch rise shape cue. This cue proved unreliable and instead key-word spotting was employed to identify such statements.

Full Paper

Bibliographic reference.  Terry, Mark / Sparks, Randall / Obenchain, Patrick (1994): "Automated query identification in English dialogue", In ICSLP-1994, 891-894.