Abstract: This paper presents a new articulatory feature based Asynchronous Dynamic Bayesian Network model（AWA-DBN） in which the dynamic pronunciation of a word is described by the moving of articulatory features.Word recognition experiments on Aurora5.0 are compared with those of WS-DBN model（in which a word is composed of a fixed number of states） and WP-DBN model（in which a word is composed of its phones）.Results show that although WS-DBN model gets the highest recognition rates，it is only suitable for small vocabulary isolated word recognition.Both AWA-DBN and WP-DBN can be adopted in large vocabulary continuous speech recognition，and AWA-DBN model gets higher recognition rates and is more robust than WP-DBN model.

Key words: Articulatory Feature（AF）, Dynamic Bayesian Network（DBN）, speech recognition

摘要： 构建了一种新的基于动态贝叶斯网络（Dynamic Bayesian Network，DBN）的异步整词-发音特征语音识别模型AWA-DBN（每个词由其发音特征的运动来描述），定义了各发音特征节点及异步检查节点的条件概率分布。在标准数字语音库Aurora5.0上的语音识别实验表明，与整词-状态DBN（WS-DBN，每个词由固定个数的整词状态构成）和整词-音素DBN（WP-DBN，每个词由其对应的音素序列构成）模型相比，WS-DBN模型虽然具有最高的识别率，但其只适用于小词汇量孤立词语音识别，AWA-DBN和WP-DBN可以为大词汇量连续语音建模，而AWA-DBN模型比WP-DBN模型具有更高的语音识别率和系统鲁棒性。

关键词: 发音特征, 动态贝叶斯网络, 语音识别