BEA is the French authority in charge of safety investigations in the event of a civil aircraft accident, and in this context analyses audio files from cockpit voice recorders (CVRs). The CVR design constraints result in the presence of a large amount of superimposed speech areas, which complicates the use of this data. In this article, we propose a CVR mixture model and present a first study on the potential contribution of Blind Source Separation for improving the intelligibility of CVR recordings.