Abstract: A synthetical pronunciation assessment technique based on multi-model matching for Mandarin initials and finals is proposed.This technique is based on single model matching and consists of three major parts：Firstly,a pronounce waveform is divided into some sub-models by searching the sonant,spirant,,ve features;Secondly,some sub-models is formed and every sub-model matching is done;Finally,the synthetical weight score is made. Experiments shew that the method of multi-model synthetical weight score has effectively improved the distinguishability to some groups of Chinese initials or finals pronounced similar.In this paper we also introduce the implementation methods of MFCC parameter extraction,DTW pattern matching algorithm,models training based on clustering,and synthetical weight score.

Key words: Speech Assessment, Pattern matching, Synthetical score

摘要： 提出了一种应用于普通话声韵母发音评价的多级音素模板综合评分法,该方法在单模板匹配的基础上,通过寻找汉语音素发音过程中的浊化、摩擦、爆破等特征,形成多个子模式,进行多模式匹配,最后给出加权评分。实验结果表明多级音素模板综合评分法有效地改善了汉语发音中几组相似音之间的区分度。也介绍了MFCC等语音特征提取、DTW模板匹配算法、基于聚类的模板训练以及综合加权评价机制的实现方法。

关键词: 语音评价, 模式匹配, 综合评分