On associative recognition of isolated Chinese word

This paper presents a new method of recognition of isolated Chinese word. Here a local feature of an input utterence to be compared with an appropriate part of every reference template.Depending on the result of local comparision(pretest), the system associates the input utterance with a sebset of recognition vocabulary, which is less than the vocabulary(set) always. Then the input utterence to be compared with the contents of the subset only, not with all templates in the vocabulary. Because the linear shift matching and a tenth of utterance length is used for ocal comparasion, the computation quantity of this stage isn´t comparable to that one for word matching by dynamic programing, therefore the new method of recognition can save much time for DP matcaing. An important problem discussed here is how to establish a subset system, which can provide a subset with appropriate size and contents for every word in the recognition vocabulary. Especially, it must be easy to establish when a vocablary or a speaker is changed. In our investigation a special algorithm was used to make the system experienced in given vocabulary and speaker automaticlly. It was based on statistics of intra-word and inter-word local distance of each word in given vocabulary nd for given speaker. A number of repetation of utterence for the vocabulary is necessary for experince mentioned above. The recognition accuracy provided by the new method was equal to that one provided by typical method using word emplates directly, it was 98.33% for the best score and 94.72% for the worst. But the time-consumption of the associative method was fourth-thirth as much as it was taken by the typical method.