of which, the shader and hal ?
not that anyone has done it yet, BUT - Im using a new software that uses these glsl codes and has them react to audio. I do it Mainly with lights, but it can be applied to a variety of parameters - proof of concept that these parameters can be changed in real time. You can encode geometric shapes with the GLSL shader ( note the pumpkin head on the shader page, but instead of a pumpkin, you make a human head ) - Hypothetically , with speech recognition you should be able to program visemes in the GLSL shader and have it talk in real time ( less any delay needed to generate the code, and render the glsl on the web page ) = there are only like 15 basic viseme shapes needed for full speech functionality.
EE, Er Ih, Ah, Oh, (w, OO), (S,Z), (Ch, J), (F, V,) TH, ( T, L, D), (B M P), ( K, G,) (N, NG,) R
the code on the link I showed you is static, but it doesn't have to be.
actually it's the same concept as a music visualiser on itunes , except that instead of flashing lights and changing colors - you focus the changes on lip shapes and visemes
this fits the bill of Roberts idea of not needing a plug in, as you don't need a plug in to view the shaders on that page, so any other person visiting your page won't need a plug in to watch it animate in real time. and the technology now exists to make it happen, but I don't think anyone has used it in that way yet - they use it as textures and music visualization. So think of it as streaming code to a web page, (java player, flash player, or html5 ) that visualizes that code through glsl.
http://learningwebgl.com/cookbook/index.php/Main_Page