I have answered similar questions before, see http://bbs.keinsci.com/forum.php ... 7&fromuid=21811 (in Chinese).
Basically, this is due to a fundamental problem with Gaussian's approach of calculating vibrationally resolved spectra: the time scales exponentially with respect to the system size, so with large molecules, Gaussian cannot finish the calculation in a reasonable amount of time, and therefore aborts.
The solution is to switch to path integral based approaches, like those implemented in ORCA or MOMAP. The computational time of these approaches scale polynomially with respect to the system size, so even very large systems can be calculated without problems.