How best to describe the asynchrony of the speech and lip motion is a key problem of audio-visual speech recognition model. A Multi-Stream Asynchrony Dynamic Bayesian Network (MS-ADBN) model is brought forward for audio-visual speech recognition, and in this model, audio stream and visual stream are synchronous in word node, while between the word nodes, each stream has its own independent phone, phone transition and observation vector node, and word transition probability is determined by audio stream and visual stream together. For each stream, each word is composed of its corresponding phones, and each phone is associated with observation feature (audio feature for audio stream and visual feature for visual stream), with some probability modeled by Gaussian mixed model. Compare with general multi-stream HMM, MS-ADBN model describes the asynchrony of audio stream and visual stream to the word level. The experiment results on continuous digit audio visual database show that: compare with multi-stream HMM, in the mismatch noise environment, an average improvement of 10.07% are obtained for MS-ADBN model.