连呼汉语识别研究

A STUDY ON RECOGNITION OF CONNECTED SPOKEN-CHINESE WORDS

本文介绍一种连呼汉语识别方法,方案中,提出一种自适应幅度规正(AAN)算法处理输入语音特征频谱,并引入声刺激量参数对其作时域非线性压缩规正。以单呼样本作参考样本。构造出一种快速动态时间弯折(FDTW)匹配过程。最后,设计出一种基于FDTW算法的二分叉树搜索决策方案进行连呼汉语识别。对十个汉语数字组成的2—5字随机序列按两种发音方式的实验给出的平均字串及字音识别率分别为:按汉语普通话发音96.8％及98.8%;按电报呼号发音97.6％及99.0%。

This paper presents a pattern matching approach to connected-Chinese word recognition. First, a method of adaptive amplitude normalization is proposed to pre-process speech data, and a sound stimulus parameter is introduced to compress and normalize the speech data. Then, Using isolated word token as reference pattern, a fast dynamic time warping (FDTW) matching procedure is perbormed. Based on the FDTW algorithm, a simplified dynamic programming decision strategy is described. Finally, the speaker-dependent recognition results for Chinese digit strings (of unknown variable length from 2-5 digts) are given with the accuracy for digit string/ digit being 96.8/98.8 percent for the standard Chin se pronunciation, and 97.6/99.0 percent for the way of pronunciation used in telecommunication.