I've recently been tossing the idea around of using quaternions to compress the tangent space data on the vertices. This is a post from my blog (look at trick #3), discussing this idea. You could potentially get away with storing only 3 quaternion components, rather than 2 sets of 2 normalized vector components. It has a roughly equivalent expense to decompress, and could be slightly faster - depending on how expensive square roots are compared with multiplies and adds.