The　image-based　lip　animation　synthesis　approach　is　one　kind　of　promising　method　that　synthesizes　the　believable　talking　head.　This　paper　seeks　to　show　an　improvement　in　the　accuracy　of　mouth　prediction　with　the　speech　stimulus,　as　well　as　showing　the　method　used　to　extract　the　speaking　mouth　correlative　speech　feature.　Our　lip　animation　synthesis　system　is　based　on　the　construction　of　a　frame　level　audiovisual　mapping　model　between　the　acoustic　speech　class　and　speaking　mouth　image　class.　Taking　the　mapping　model　as　a　basis,　genetic　algorithm　is　used　to　extract　the　speaking　mouth　correlative　speech　feature.　The　key　step　used　in　this　study　is:　fitness　and　coding　scheme　designing.　Experimental　results　show　that　the　extracted　speech　feature　has　a　better　correlation　with　the　corresponding　speaking　mouth,　compared　to　the　single　or　mixed　LPCC　and　MFCC.　More　research　will　be　done　in　this　specialist　field　of　study　the　multi-layer　speaking　mouth　correlative　speech　feature　abstraction　structure,　and　will　attempt　to　show　that　the　speaking　mouth　correlative　speech　feature　should　have　better　results.　©　2006　IEEE.