In this paper, we consider a specific part of statistical machine translation: feature estimation for the translation model. The classical way to estimate these features is based on relative frequencies. In this new approach, we propose to use the concept of belief masses to estimate the phrase translation probabilities. The Belief Function theory has proven to be suitable and adapted for dealing with uncertainties in many domains. We have performed a series of experiments to translate from English into French and from Arabic into English showing that our approach performs, at least as well as and at times better than, the classical approach.