构建一种基于发音特征的音视频双流动态贝叶斯网络(DBN)语音识别模型(AF_AV_DBN),定义节点的条件概率关系,使发音特征状态的变化可以异步。在音视频语音数据库上的语音识别实验表明,通过调整发音特征之间的异步约束,AF_AV_DBN模型能得到比基于状态的同步和异步DBN模型以及音频单流模型更高的识别率,对噪声也具有较好的鲁棒性。 This paper constructs a speech and audio dual-stream Dynamic Bayesian Network (AFN) speech recognition model (AF_AV_DBN) based on pronunciation features, and defines the conditional probability relationships of nodes so that the variation of pronunciation feature states can be asynchronous. Speech recognition experiments on audiovisual and voice databases show that the AF_AV_DBN model can achieve higher recognition rate than the state-based synchronous and asynchronous DBN model and audio single-stream model by adjusting the asynchronous constraints between pronunciation features, Better robustness.