Speaker-independent word recognition using dynamic programming matching with statistic time warping cost

A speaker-independent word recognition method is presented to improve word recognition performance without using a large training pattern set. The method is based on a multiple-template technique for treating spectral variations, combined with a statistical approach for temporal variations treatment. Dynamic programming is used for matching without local path constraint into order to fully normalize the temporal variations, and statistical time warping cost (path cost) is attached to each template to penalize unnatural time warping. The path cost is derived from the path distribution, which is obtained from a set of time-warping functions set between individual training patterns in the cluster and the cluster center pattern. Evaluation experiments were conducted, using 23-word vocabulary samples uttered by 107 speakers. The correct recognition rate was 98.8% with path cost (two templates/word) and 96.3% without path cost (four templates/word)