论文部分内容阅读
本文描述了一种具有时间紧缩算法的汉语连续字识别系统。在文献[1]工作的基础上,改进了端点检测和字串分割的方法,并利用汉语语音短时平均零交叉率的统计特性,对被识别字的时宽实行紧缩处理。实验结果表明,经过这种处理之后,不仅使参考模式与试验模式的时间模型得到基本对准;而且平均紧缩率可达64％,为减少模式特征参数的计算量和存储量,实现大词汇表识别,提供了有利条件。实验还表明,通过上述处理,识别率仍可达到94％左右。
This paper describes a Chinese word recognition system with time squeezing algorithm. Based on the work in [1], the method of endpoint detection and string segmentation is improved. By using the statistical characteristics of short-time average zero-crossing rate of Chinese speech, the duration of the identified words is tightened. The experimental results show that not only the time model of the reference mode and the experimental mode are basically aligned, but also the average compression ratio can reach 64%. In order to reduce the computation and storage of the mode characteristic parameters, a large vocabulary Identify and provide favorable conditions. Experiments also show that through the above treatment, the recognition rate can still reach about 94%.