Segmentation for data reduction in isolated word recognition

Dynamic programming has been shown to give excellent results for isolated word recognition. Two major drawbacks to dynamic time warping are the excessive storage and computational requirements for large vocabularies. This paper proposes a technique of data reduction using acoustic segmentation which gives savings in both these areas. To demonstrate the advantages of this technique three distance measures are evaluated for performance in the unsegmented case. The one yielding the best recognition accuracy is used to compare three algorithms for matching segmented templates. Performance in terms of recognition accuracy, computation time, and savings in storage are given.