Replaced calls to np.tile with ndarray broadcasting (uses less memory,
moves more work to C library).  This sped up HMM training on a large
database (chroma from 3000 songs) by about 33%.