基于等级式多层感知器网络的特定人汉语词组识别系统

SPEAKER-DEPENDENT RECOGNITION SYSTEM OF CHI-NESE PHRASES BASED ON HIERARCHICALLY MULTILAYERED PERCEPTRON(MLP)

本文主要讨论多层感知器网络￣［１，２］在特定人的汉语词组识别系统中的应用。提出简单的语音特征参数规整方法。针对多层感知器网络在分类模式增加时，网络结构增大，学习时间冗长，识别率下降；增加新的模式需要重新训练网络的缺点，提出了等级式多层感知器网络及其新的分类算法。基于语音知识建立了一个中等词汇量的特定人的汉语词组识别系统，提出了语音知识的矩阵表达方式，基于前向概率和后向概率的最佳匹配推理算法。该系统的音节识别率为９０％，根据知识推理，词组的识别率为９３．５％。

In the paper we develop the speaker-dependent recognition system of Chinesephrases based on hierarchically multi-layered perceptron. The system is composed of the following two main parts:(1)syllable recognition;(2)phrase recognition based on knowledge. A hierarchically multi-layered perceptron(HMLP)is proposed to identify syllable.The HMLP is composed of type classification and pattern recognition networks,the formerclassifing input patterns into different types and the latter recognizing input patterns.A Chinese syllable connection rule is presented,the forward and the backward connection informations between the syllables is applied to recognize phrases,The experiment results show thatthe system is very effective. The syllable recognition accuracy is 90%and the phrase recognition accuracy is 93.5%。