加快图样匹配过程的一条途径

本文提出一种加快图样匹配过程的新途径。它将原来的数字式图样分成四个子图样,分四步进行匹配比较,在每一步上进行一个子图样的匹配比较,选出一部分跟未知图样距离较小的参考图样,参加下一步的图样匹配比较,其余淘汰。这样使匹配过程加快三倍以上。文中给出了在语言识别方面的实验结果。采用二种淘汰办法,它们在前三级上的淘汰率分别是:（i）3/4,1/2,1/2;（ii）9/10,1/2,1/2。这二种办法所需的图样匹配时间分别是0.359 T₀和0.294 T₀,T₀——是完整的匹配过程所需的时间,而正确率几乎不下降。这样我们原来的通用实时语言识别系统RTSRS（01）就能实时识别600条口令。

AN APPROACH TO SPEED THE PROCESS OF PATTERN MATCHING

A new approach for speeding the process of pattern matching is presented. The digital pattern （unknown or reference） is divided into 4 （or more） subpatterns of equal size, and the pattern matching is performed in 4 steps. At every step only one subpattern is used to perform pattern matching, and a part of reference patterns which have less distances to the unknown pattern are retained to match further in the next step. In this approach, the pattern matching can be speeded 3 times or more faster than the original one. A set of experiment results rate on speech recognition using such approach is given. It is shown that the correct recognition obtained by the enw method doesn't descend apparently when compared with the results of the complete pattern matching. In this way our real-time speech recognition sys-tem-RTSRS （01） can recognize 600 spoken commands in real-time.