This paper presents a corpus design method for Text-To-Speech (TTS) synthesis application. The aim of this method is to build a corpus whose unit distribution approximates a given target distribution. Corpus selection can be expressed as a set covering problem, which is known to be NP-complete: we therefore resort to a heuristic approach, based on greedy algorithm. We propose the Kullback-Leibler divergence to guide the iterative selection of candidate sentences: indeed, this criterion gives the possibility to control the unit distribution at each step of the algorithm. We first show how to efficiently update, in an incremental manner, this criterion. We then present and discuss experimental results, where our selection algorithm is compared, for various unit sets, with alternative selection criteria.