APPLICATION OF CLUSTERING TECHNIQUES TO SPEAKER-TRAINED ISOLATED WORD RECOGNITION

Speaker‐trained, isolated word recognizers have achieved notable success in a wide variety of applications. The training for such systems generally involves a single (or sometimes two) replication(s) of each word of the vocabulary by the designated talker. Word reference templates are then formed directly from these replications. In recent work on speaker‐independent word recognition, it has been shown that statistical clustering procedures provided an effective way for determining the structure in multiple replications of a word by different talkers. Such techniques were then used to provide a set of reference templates based on the clustering results. In this paper we discuss the application of clustering techniques to speaker‐trained word recognizers. It is shown that significant improvements in recognition accuracy are obtained when using templates obtained from a clustering analysis of multiple replications of a word by the designated talker. It is also shown that recognition accuracy did not change with time (over a 6‐month period) for any of the subjects tested, thereby indicating that the reference templates were reasonably stable. © 1979 The Bell System Technical Journal