Recently I discovered Armadillo
Is licensed as LGPL, also based on LAPACK, MKL or ACML, with similar basic math. In benchmarks achieved high speedups against IT++, Octave and MATLAB. Some benchmarks you will find here:
http://arma.sourceforge.net/speed.html
I didn't fully understand the reason, but they claim to use "lazy evaluation" and therefore need less temporary math variables internally.
However, Armadillo today is somehow limited to simple vector/matrix math. It does not have the feature-rich library of signal processing functions like IT++.
Now what about this idea: mix the two libraries in a new fork, taking the best of the two worlds and create a new library using the Armadillo math as basics and adapt the signal processing blocks of IT++. I think most of the signal processing code does not have to be changed, just replacing the variable type for vectors and matrices.