In this article we aim to present the adaptation of an automatic speech recognition system for specific and robust applications related to the fields of medicine and disability. It constitutes the first step towards the building of an open source system designed to automatically translate speech into pictograms for allophone speakers or people having cognitive disorders in emergency medical settings. Due to the criticality of an adequate speech recognition in such contexts, we decided to rely on formal grammars representing the natural language used. Some preliminary experiments are displayed in order to evaluate its use in real-life situations.