An connected word recognition algorithm based on the removable rule of Chinese acoustical-tail and hypothesis-verification perceptual theory

In this paper, a removable property of acoustical-tail of Chinese word is described. And 3 conclusions can be drawn as the removable rule:First, if the part of the acoustical-tail's feature vectors are suitably removed, the recognition rate will not decrease considerably, in contrast, it even can increase. Second, the recognition time abbreviates obviously without a part of acoustical-tail's feature vectors. Third, it is also indicated that the limit of removal of acoustical-tail's feature vectors should be no more than 1/3 of all the feature vectors. Applying the removable rule to the isolated word recognition, a fast algorithm is proposed. Furthermore, on the,view of Hypothesis——Verification perceptual theory, a connected word recognition algorithm is successfully implemented with the model of DTW. In this algorithm, there is no limit to the length of the connected vocabulary to be recognized, and the recognition time only increase linearly with the increasement of the length of the vocabulary. In the experiment of 200 connected vocabulary of speaker-dependent (including 162 two-word, 22 of vocabulary there-word and 16 of four-word), a recognition rate of 91% is obtained.