We present a novel speech animation approach using coupled hidden Markov models (CHMMs). Different from the conventional HMMs that use a single state chain to model the audio-visual speech with tight inter-modal synchronization, we use the CHMMs to model the asynchrony, different discriminative abilities, and temporal coupling between the audio speech and the visual speech, which are important factors for animations looking natural. Based on the audio-visual CHMMs, visual animation parameters are predicted from audio through an EM-based audio to visual conversion algorithm. Experiments on the JEWEL AV database show that compared with the conventional HMMs, the CHMMs can output visual parameters that are much closer to the actual ones. Explicit modelling of audio-visual speech is promising in speech animation.