In the context of a spoken interfaces, we present a methodology and an implemented workbench called EGAL (Lexicalized Tree Grammar Extraction) dedicated to design and test restricted languages used in specific task-oriented applications. A complementary methodology is proposed to process the extraction of these applicative languages from a general LTAG grammar and a training corpus. Additional results allow us to estimate the representativeness of the training corpus. An application of the system is presented for the tuning of a LTAG grammar dedicated to a spoken interface on the basis of a Wizard of Oz corpus. || Nous présentons une méthododologie pratique et une plate-forme appelée EGAL (Extraction de Grammaire d'Arbres Lexicalisés) dédiés à la conception et au test d'un sous-langage d'application pour un système de dialogue oral homme-machine. Une méthodologie c