ABSTRACT
Given the present state of automatic speech recognition (ASR) technology, it is clear that the design of error tolerant systems is an important area of study. Furthermore, current technology does not allow for fully automatic detection and correction of errors, so there exists a need for error correction dialogues which remain under human control. While errors can arise in both the device and the user (see Baber et al., 1990), in this paper we assume that user errors have been minimised in order to focus our attention on recognition errors. Of the three categories into which recognition error are generally classified, viz. insertions, classifications, substitutions, the most common form is substitutions (Brown and Vosburgh, 1989). Not only are errors arising from substitutions the most common, but also these present the biggest problem in terms of prevention and correction. This type of error is, therefore, the main focus of efforts to develop error detection and correction strategies.
