Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Improving Recognizer Acceptance Through Robust, Natural Speech Repair

Arthur E. McNair, Alex Waibel

Carnegie Mellon University, Pittsburgh, PA, USA

Though large vocabulary speech recognition systems have improved greatly in recent years, usability of these systems in practical applications is still low, due to the ever-present errors. Adding a natural interface to these systems for users to correct errors should increase acceptance. This paper describes three methods for accomplishing speech-based repair of a misrecognition. The user must respeak or spell only an errorful subsection of the original utterance. A method is described to automatically locate the respoken subpiece in up to 90% of the instances. Once the subpiece has been respoken and located, another method is described which corrects the subpiece in up to 70% of the instances. If the location is known, and the subpiece is spelled, a third method is described which uses a single spelling utterance to correct the subpiece in up to 82% of the instances. The results indicate that these methods can decrease the error rate of a CSR by two thirds using only a single short repair utterance.

Full Paper

Bibliographic reference.  McNair, Arthur E. / Waibel, Alex (1994): "Improving recognizer acceptance through robust, natural speech repair", In ICSLP-1994, 1299-1302.