The spectral-based characterization of inkjet printers is often based on a physical description of the printing process. The objective of our work is to see whether an approach based on the use of neural networks is an effective strategy for spectral printer characterization without requiring a deep knowledge of the printing process. In our experiments, we treat the printers as RGB devices, and exploit finite-dimensional linear models to reduce the amount of information required to characterize them. To select a good architecture, we compared the behavior of 15 different networks to compute reflectance spectra from RGB digital counts. To test our characterization procedure we consider an Epson 890 inkjet printer using photo quality paper.