Order dependence in templates for monosyllabic word identification

An efficient algorithm using dynamic programming for temporal registration of reference words to form templates is potentially hindered by order dependence. A different template will result for each ordering of reference words. For a large number of words, the ordering is probably not significant. However, for a small number, ordering may be significant. For each of two speakers, ten sets of fifty-two CVC word templates based on mel-frequency cepstral coefficients were generated. Each template comprised three to five randomly-ordered words. The average correct identification in open tests for each speaker was 94.76% and 90.53%, with standard deviations of 0.53% and 0.48%, respectively. Thus the ordering of words during template generation did not significantly affect word identification.