I'm experimenting with this at the moment, and my configuration is to store the normal and tangent in 16-bits each (using schemes from this page), two roughness parameters in 8 bits each, and the sign of the binormal (whether it's cross(N,T) or -cross(N,T)) in one bit.
This requires 49 bits of G-buffer space, rather than 24 bits for isotropic specular...
I also tried storing the T/B/N basis as a quaternion, but to get decent quality I needed to use 16-bits for each of the 4 components. I was able to drop one component and reconstruct it (due to quaternion rotations being normalized), bringing it down to 16-bits * 3, but this was still way more space and conversion math than the above configuration.
Crytek mention in one of their presentations (forget which one, sorry) that they faked anisotropic specular by simply adjusting the normal when writing to the g-buffer.