Boron carbide, a lightweight, high temperature material, has various applications as a structural material and as a neutron absorber. The large solubility range of carbon in boron, between $sim$ 9 % and 20 %, has been theoretically explained by some of us by the thermodynamical stability of three icosahedral phases at low temperature, with respective carbon atomic concentrations: 8.7 % (B$_{10.5}$C, named OPO$_1$), 13.0 % (B$_{6.7}$C, named OPO$_2$), whose theoretical Raman spectra are still unknown, and 20 % (B4C), from which the nature of some of the Raman peaks are still debated. We report theoretical and experimental results of the first order, non-resonant, Raman spectrum of boron carbide. Density functional perturbation theory enables us to obtain the Raman spectra of the OPO$_1$ and OPO$_2$ phases, which are perfectly ordered structures with however a complex crystalline motif of 414 atoms, due to charge compensation effects. Moreover, for the carbon-rich B$_4$C, with a simpler 15-atom unit cell, we study the influence of the low energy point defects and of their concentrations on the Raman spectrum, in connection with experiments, thus providing insights into the sensitivity of experimental spectra to sample preparation, experimental conditions and setup. In particular, this enables us to propose a new structure at 19.2 % atomic carbon concentration, B$_{4.2}$C, that, within the local density approximation of density functional theory (DFT-LDA), lies very close to the convex hull of boron carbide, on the carbon-rich side. This new phase, derived from what we name the "3+1" defect complex, helps in reconciling the experimentally observed Raman spectrum with the theory around 1000 cm-1. Finally, we predict the intensity variations induced by the experimental geometry and quantitavely assess the localisation of bulk and defect vibrational modes and their character, with an analysis of "chain" and "icosahedral" modes.