抄録
The effective reference pattern set structure and the training procedure for Japanese syllable recognition are reported. The studies of multi-reference pattern and training procedure were conducted for the optimum reference pattern structure to cover wide utterance variations. The examined training procedure included the sequential training to compensate the time dependent variation of pronounced pattern. The experiments proved that the recognition accuracy depends on the number of training rather than the number of reference pattern set, and also that the aggravation of recognition accuracy caused by the varied utterance with time is improved by the sequential training conducted after every recognition step.
