BabelDr is a medical speech to speech translator, where the doctor has to approve the sentence that will be translated for the patient before translation; this step is done using monolingual backtranslation, which converts the speech recognition result into a core sentence. In this work, we model this step as a simplification task and pro-pose to use neural networks to perform the backtranslation by generating and choosing the best core sentence. Results of a task-based evaluation show that neural net-works outperform previous versions of the system.