Semi-empirical quantum chemistry methods offer a very interesting compromise between accuracy and computational load. In order to assess the performance of NDDO methods in the interpretation of terahertz spectra, the low frequency vibration modes of three crystalline materials, namely, polyethylene, poly(vinylidene fluoride) form II and α-D-glucose have been studied using the PM6 and PM7 Hamiltonians and the results have been compared with the experimental data and former calculations. The results show good qualitative or semi-quantitative agreement with the experimentally observed terahertz spectra. arXiv:1604.03919v1 [physics.chem-ph] 13 Apr 2016 to permit the identification of the transition from a 2D form to a 3D arrangement in a carbon nitride polymer [3]. The absorption in the THz has been demonstrated to display a direct correlation with the stacking order along the c axis in different types of carbon materials [4].The complexity of the underlying physical effects affecting the observed attenuation, with contributions from intra-and inter-molecular modes and a strong impact of disorder, hinders the interpretation of the data. Theoretical methods are essential for this task, and the versatility of computational quantum chemistry methods is particularly attractive for the discernment of THz spectra.The theoretical prediction of the terahertz spectra of crystalline materials has been based mostly on force-field or density functional theory (DFT) methods [1]. Ab-initio and density functional theory (DFT) methods are regarded as highly accurate in the prediction of molecular vibrations [5] when used with correction factors to compensate systematic deviations. Nevertheless, the relevance of long-range and intermolecular contributions can make the study of the THz vibrational spectra of materials very demanding in terms of the required solid-state calculations over large systems. This imposes a severe limitation in the use of computationally costly ab-initio or DFT methods. On the other hand, semi-empirical methods [6,7], specially with the recently developed parallel implementations for shared-memory multiprocessor and massively-parallel graphics processing units architectures [8], permit to study the vibrations in large molecular systems and crystals at reduced computation times.Semi-empirical NDDO (neglect of the diatomic differential overlap) computations are based on approximations and data obtained from given training sets which calls for their assessment when employed in a given task, and can be placed at an intermediate theory level between force-field and ab-initio or DFT methods. Poor performance of the AM1 [9] and the PM3 [10] methods reported in [6] disfavored their use in the analysis of vibrational spectra until the appearance of the PM6 parametrization [11], with improved accuracy in the prediction of geometries and the description of the material vibrations [7]. The recently developed PM7 [12] method yields higher accuracy in the guessed structures, particularly for large molecular crystals, even ...