Modelling Error Recovery and Repair in Automatic Speech Recognition

Abstract While automatic speech recognition (ASR) has achieved some level of success, it often fails to live up to its hype. One of the principal reasons for this apparent failure is the prevalence of "recognition errors". This makes error correction a topic of increasing importance to ASR system development, with a growing awareness that, by designing for error, a number of problems can be overcome. Currently, there is a wide range of possible techniques which could be used for correcting recognition errors, and it is often difficult to compare the techniques objectively because their performance is closely related to their implementation. Furthermore, different techniques may be more suited to different applications and domains. It would be useful to have some means of defining the requirements of an error correction dialogue, based on characteristics of the dialogue and ASR system in which it is to be used, in order to develop design specifications for appropriate error correction. This paper reports an approach, based on task-network modelling, which could be used to this end.