In　order　to　generate　more　realistic　mouth　animation　in　visual　speech　synthesis,　this　paper　proposed　a　method　based　on　a　two-level　learning　model.　The　authors　can　learn　the　potential　mapping　relationship　between　acoustic　features　and　the　visual　features　through　the　combination　of　HMM　(Hidden　Markov　Models)　and　GA　(Genetic　Algorithms).　This　model　can　decrease　the　redundant　information　in　abstracting　acoustic　features　for　large　acoustic　sample　space　and　predict　more　realistic　mouth　animation.　In　addition,　this　paper　also　proposed　a　new　method　based　on　FAP　points　in　mouth　feature　expression.　This　method　can　eliminate　the　effect　by　illumination　and　decrease　the　dimensions　of　mouth　feature　vector.　It　improves　the　speed　of　training　and　synthesis.