Particle size fractionation enables a better understanding of soil organic carbon (C) dynamics since it separates fractions that differ in composition, residence time and function. However, this method is time-consuming and tedious; thus, its use has been greatly limited. Our objective was to evaluate the ability of an existing soil spectral library (SSL) from different regions of West Africa to predict the C amount in the fractions (gC kg-1 soil) of the samples in a new target set from Benin. The SSL included 181 samples from five countries, and the target set included 94 samples (depth = 40 cm), most of which were coarse-textured; near-infrared reflectance (NIR) spectra were collected for 2 mm sieved samples (non-fractionated samples). The predicted variables were the C amounts in the non-fractionated soil and in the < 20, 20–50, and > 50 µm fractions (F<20, F20–50, and F>50, respectively). Different methods were tested to optimize the predictions: (i) SSL enrichment with 10 or 15 samples selected from the target set (spiking) and replicated six times (i.e. extra-weighted); (ii) locally weighted (local) partial least squares regression (PLSR), which is calibration by the spectral neighbours with the highest weights attributed to closest neighbours, and was compared to “global” (i.e., common) PLSR, where all calibration samples equally contribute; and (iii) spectrum pretreatments (e.g., smoothing, centring, derivatization). In addition, the intermediate precision of the conventional data (standard error of laboratory; SELint) was estimated through triplicate fractionation of three samples carried out by three operators (one per replicate). When the SSL alone was used for calibration, the predictions were inaccurate for the C amounts in the non-fractionated soil and in F<20; however, the predictions were accurate for the C amounts in F20-50 and F>50, with minimal benefit from the local PLSR over the global PLSR in general. For the non-fractionated soil, F<20, F20–50