Formation of soot in an auto-igniting n-dodecane spray under diesel engine relevant conditions has been investigated numerically. The bulk of research thus far has addressed turbulence-chemistry interaction by coupling highly diffusive turbulence models with more sophisticated combustion models. Instead, this study employs the advanced sub-grid scale k-equation model in the framework of a Large Eddy Simulation (LES) together with the uninvolved Direct Integration approach. A reduced n-heptane chemical mechanism has been employed and artificially accelerated in order to predict the ignition for n-dodecane accurately. Soot processes have been modelled with an extended version of the semi-empirical, two-equation model of Leung, which considers C2H2 as the soot precursor and accounts for particle inception, surface growth by C2H2 addition, oxidation by O2, oxidation by OH and particle coagulation. A full spray event has been simulated and statistics have been collected over the quasi-steady state of the spray. The results are compared to experimental data from the Engine Combustion Network in terms of vapor penetration, lift-off length, global soot mass and distribution of soot volume fraction. They are also compared to Reynolds-averaged Navier-Stokes (RANS) computations in terms of time-averaged fields related to flame structure and soot kinetics. Both turbulence modelling frameworks are shown to successfully predict the time-averaged distribution of soot volume fraction in physical space. However, LES is superior to RANS, because it additionally succeeds at capturing the intermittency of soot formation and oxidation. Comparison between terms in the soot mass fraction transport equation provides insight into the origin of the qualitative differences between the two frameworks.