Error detection and correction in a speech recognition system: a knowledge based system approach
Kyung-ho Loken-Kim · 1988
The purpose of this study was to explore the possibility of applying high-level knowledge sources, such as; syntax, semantics, and error history, to automatically detect and correct recognition errors generated by a Vocalink Model CSRB speaker dependent speech recognizer. A pilot study was conducted to understand the behavior of the speech recognizer, and the results are presented here. The result of the pilot study led to the development of AUTODAC (Automatic Error Detection and Correction System). The characteristics of AUTODAC are: (1) It is a blackboard based system. (2) It can parse ill-formed sentences with a combination of left-to-right and right-to-left parsing. (3) It can learn the history of recognition errors and utilize this information to subsequently recover similar recognition errors later. (4) It allows a user to manually correct any part of the recognized sentence. A flexible manufacturing center was selected as the task domain. A set of 12 sentences that are presumed to be commonly used in the task environment were composed. Each of the 12 sentences was repeated 16 times, thus, a total of 192 sentences were tested to evaluate the performance of AUTODAC. The results show that a total of 142 out of 192 sentences tested were recovered with the combination of automatic and manual error correction. The results of this study show that the use of high-level knowledge sources for the purpose of controlling recognition errors appears promising.