Method for determining an acoustic model for a word

摘要

The method involves deviating characteristics (11,12,13,21,22,23,31) from speech signals of different words in a training phase to repeated time points. The characteristics identify consecutive states in triphons in the words. The triphon consists of a left, a central and a right phoneme. Characteristics of different triphons which have the same acoustic states and satisfy predetermined criterions are combined to groups. A group of triphons is selected for each state of a triphon in the first word which was not observed in the training phase. The selected group is assigned to observed triphons having the same central, left and right phonemes. Each selected group formes a reference value for a word model.