Abstract:
KDD has developed a telephone operator assisting system, which provides operators with automatic database inquiry using speech recognition. An outline of the system is described and the results of a field trial are evaluated. There have been many speech recognition systems for practical use. A fast recognition method is one of the important subjects for practical systems. For this purpose, various methods are proposed. [Sagayama et al. ``New Approaches Toward Fast Speech Recognition,'' Proc. Acoust. Soc. Jpn. 25--28 (1996) (in Japanese)]. However, it takes some frames to output recognition results after finishing speech input, because any method deals with a few 100-ms pause following speech, which is a margin to detect the speech end. Therefore, a new fast recognition method is proposed. The basic idea of this approach is to output results when inquiry keywords are decided, even before detecting the speech end. This approach makes it possible not only to reduce speech recognition time, but also to recognize speech including out-of-vocabulary words and nonspeech sounds at the end of a sentence. Evaluation results show that (1) speech recognition time was reduced 30% and (2) high recognition accuracy was achieved: 95% for only expected inputs, 92% for all inputs. Fifty percent of unexpected inputs included out-of-vocabulary words at the end of a sentence, and most of them can be recognized correctly using this method.