This study is based on the exploitation of telephone calls to hospital emergency departments in order to train artificial intelligence tools for the automatic recognition of vocal emotions for the purpose of improving emergency medical regulation. The main objective is to extract the most characteristic acoustic cues of vocal emotions. Very limited samples were selected from the audio databases of two University Hospitals in the Jura Arc (CHRU of Besançon in France and CHUV of Lausanne in Switzerland). Only extracts containing linguistic vocal emotions with a negative polarity (angst, anger, embarrassment and sadness) were retained for the study. These extracts were segmented and labeled, and their acoustic parameters (F0, intensity and duration) were measured manually. The preliminary results show that F0 modulations seem more accurate to discriminate the four vocal emotions.