In this study, we aimed at determining statistical models that allowed for the classification of impact sounds according to the perceived material (Wood, Metal and Glass). For that purpose, everyday life sounds were recorded, analyzed and resynthesized to insure the generation of realistic sounds. Listening tests were conducted to define sets of typical sounds of each material category by using a statistical approach. For the construction of statistical models, acoustic descriptors known to be relevant for timbre perception and for material identification were investigated. These models were calibrated and validated using a binary logistic regression method. A discussion about the applications of these results in the context of sound synthesis concludes the article.