摘要
本文指出，如果适当切除单字部分音尾特征，识别率不会明显下降，甚至有所提高，而识别时间明显缩短。分析和实验结果表明：若采用动态时间规正算法（ＤＴＷ），识别时间与特征矢量长度的平方成正比关系。本文对上述现象进行了探讨，并给出了直观解释。实验指出，音尾特征的截除极限为特征矢量总长度的１／３。根据上述实验事实，从假设－检验的认知理论出发，提出一种汉语连接词的识别算法，并在ＤＴＷ模型上得以实现。实验测试集包括２００个特定人发音样本，其中２字词１６２个，３字词２２个，４字词１６个，正确识别率为９１％。该算法对待识词的长度没有限制，并且随待识长度的增加，识别时间只作线性增长。
It is indicated that the recognition rate of isolated Chinese words will not decrease considerably, if the part of the phonetic ends′ feature vectors are suitably removed. In contrast, it may even be increased. At the same time, the recognition time will be significantly reduced. Both analysis and experiments show that the recognition time is proportional to the square of the length of the feature vector,if dynamic time warping (DTW) algorithm is employed in recognition. An explanation is given after the previous phenomenon has been discussed. It is also pointed out that the utmost of removal of phonetic ends′ feature vectors should be no more than 1/3 of all the feature vectors. According to the statement of removable rule, connected word recognition algorithm is submitted in view of the Hypothesis Verification perceptual theory, this algorithm is successfully implemented with the model of DTW. In the experiment of 200 speaker dependent phonetic samples, including 162 of 2 Chinese characters vocabulary, 22 of 3 Chinese characters and 16 of 4 Chinese characters ones, a recognition rate of 91% is obtained. In this algorithm, there is no limit to the length of syllables to be recognized. Furthermore, the recognition time increase only linearly with the increase of the phonetic length.
出处
《数据采集与处理》
EI
CSCD
1998年第4期311-314,共4页
Journal of Data Acquisition and Processing
关键词
言语识别
语言信号处理
连接词识别
算法
speech recognition
speech signal processing
connected word recognition
dynamic time warping
